Agora Custom Development Services: 2026 Cost Guide

Why Fora Soft wrote this Agora playbook

Fora Soft is a software development company that has built video, voice and real-time communication products since 2005 — 250+ shipped products and 50 in-house engineers. Agora, Inc. (agora.io, NASDAQ: API) is the real-time engagement platform whose SDKs those products often run on: video, voice, live streaming, signaling, chat and AI voice agents delivered over its own private edge network, which Agora calls SD-RTN, the Software Defined Real-Time Network. This is not a review. It is the scoping conversation we have every few weeks, written down, with the 2026 numbers attached.

We have built on Agora specifically since 2011 and are listed in the Agora partner directory. On that network we have taken real-time video products from ten users to 100,000 concurrent participants, and our embeddable Agora classroom component runs on Agora Flexible Classroom — video, whiteboard, document editing, screen share and collaborative playback, deployable into an existing EdTech product in about a week. We have also shipped the opposite call: BrainCert runs its own WebRTC stack, CirrusMED is HIPAA telemedicine, and Perspire.tv runs LiveKit. Which platform a product lands on is a decision, not a default, and we would rather tell you “use LiveKit” on a scoping call than bill you for the wrong architecture.

Three things here correct what almost everyone else publishes, including an earlier version of this page. Agora bills video on aggregate subscribed resolution, so room size drives your rate more than your encoder settings do. The Conversational AI Engine costs $0.10 per minute, not the $0.0265 figure that circulated through 2025. And Twilio never killed Programmable Video — it reversed that decision in October 2024. All three change which platform you should pick and what you should budget.

What custom Agora development services actually cover

Custom Agora development services buy you seven things that sit around the call: a token service, a recording orchestrator with retention and redaction rules, a moderation pipeline, an AI orchestrator, a multi-tenant model, a quality-of-experience and cost-metering layer, and the compliance wiring that makes an auditor sign off. They do not buy you the calling SDK — a senior engineer can paste an App ID into a sample app and have two people on a video call the same afternoon.

This article prices and specifies that work. If what you actually want is team composition, rates and start dates, go to our Agora development team page instead — that is where the engagement models live.

That is the honest answer to “how much does custom Agora development cost”: the SDK integration is a rounding error, and everything above it is eight to sixteen weeks of senior work. The rest of this playbook prices each of those layers, shows the arithmetic on a real workload, and gives you the five questions that decide whether you need any of it.

Agora pricing in 2026: the full line-item list

Agora publishes every price, which is more than most CPaaS vendors do — but it publishes them across a dozen separate pages, so nobody sees the whole bill at once. Here is the whole bill, pulled from Agora’s own pricing pages on 18 July 2026.

Figure 1. Twelve Agora SKUs a custom build actually gets billed for, plus the Cloud Proxy minimum monthly fee that never appears in comparison articles.

The billing unit matters more than any single rate, and it works in two steps. First, Agora bills per participant-minute, not per session-minute: a 20-minute two-person consult is 40 billable minutes, and a 20-minute class with an instructor and 30 students is 620. Second — and this is the step that surprises people — those minutes are converted into Standard minutes at a ratio set by the video tier, and the tier is set by the total resolution each user subscribes to, not the resolution anyone publishes: audio 1:1, HD 1:4, Full HD 1:9, 2K 1:16, 2K+ 1:36. Everything then bills at $0.99 per 1,000 Standard minutes. The list prices above are those ratios rounded up: 4 × $0.99 is $3.96 and the published HD rate is $3.99.

Two things partly offset this, and both have a catch. Every account — not every project — gets 10,000 free Standard minutes a month, and Agora is explicit that those free minutes stop applying the moment you buy a paid package. Prepaid packages are the discount mechanism instead: $1,217.99 buys 1.5 million Standard minutes, an 18% saving against the $0.99 audio rate. Read that number carefully. At the 1:4 HD ratio, 1.5 million Standard minutes is 375,000 HD video participant-minutes, not 1.5 million of them. Neither the package discount nor the top-up discount touches the Cloud Proxy base fee.

One more mechanic worth knowing before you model anything: if a user sits in a channel and neither publishes nor subscribes to a thing, Agora still charges them the audio rate for the time they were connected. Empty waiting rooms are not free.

Four line items that quietly bleed Agora budgets

The four line items that most often blow up an Agora bill are aggregate subscribed resolution, recording priced on that same aggregate, transcription scope nobody set, and Cloud Proxy. Together they are usually the gap between what the team modelled — per-minute rate times session minutes — and what the invoice says.

1. Aggregate resolution, which is the one that actually does the damage. Agora does not bill you for the resolution you publish. It bills each user for the sum of the resolutions they subscribe to, and that sum picks the tier. One 1280×720 subscription is 921,600 pixels, which is exactly the HD ceiling. Add a second person to the room and each user is now subscribing to two streams, 1,843,200 pixels, and everyone bills at Full HD. At four people it is 2K. At six it is 2K+. Same cameras, same encoder settings, nine times the rate per participant-minute.

Figure 2. Every stream here is 1280×720. Only the number of people changes, and the billed tier climbs with it.

Screen share is the express lane into the top tier: the default capture resolution is 1920×1080 on Windows, macOS and Web, and those 2,073,600 pixels land in every subscriber’s aggregate at once. Agora’s own worked example has two audience members in a five-person session billing at 2K+ purely because a host shared a screen. The lever that actually works is the role model, not the encoder: passive viewers moved to the Broadcast Streaming audience role convert an HD minute at 1:2 instead of 1:4, and dual-stream mode bills subscribers on the small stream they actually receive.

2. Recording priced on aggregate too. A useful correction to something we see repeated everywhere, including in our own earlier draft: the recording fee does not depend on whether you pick composite or individual mode, and per-stream minutes are not additive. Agora prices recording on the number of streams recorded, the duration, and the aggregate recorded resolution — the same tier mechanic as the call itself. So recording a five-host channel does not cost five times a one-host channel; it costs whatever tier the combined resolution lands in. Pick the mode from your redaction model instead: composite is one file and cannot be un-mixed, individual makes a deletion request a file delete.

3. Transcription scope nobody set. At $16.99 per 1,000 minutes, Real-Time Speech to Text is the most expensive thing you can leave switched on by accident — it costs more than four times Video HD. Agora is fairer here than it gets credit for: minutes where every host is muted are excluded from the transcription rate and fall back to a $0.99 standby fee. That protects you from silent participants; it does nothing about a channel where somebody is talking and nobody needs the transcript. A 300,000-minute month of transcribed audio is 300,000 ÷ 1,000 × $16.99 = $5,097, and on the platforms we audit a large share of that is waiting-room chatter, hold queues and small talk before the real conversation starts. Decide per session type what deserves a transcript, and gate it through the SDK control hooks rather than enabling it channel-wide.

4. The Cloud Proxy floor. This is the one nobody writes about. Cloud Proxy is Agora’s feature for reaching users whose network blocks ordinary media traffic: if your users sit behind hospital, bank or school firewalls, you enable it in Force UDP or Force TCP/TLS 443 mode. Media minutes then carry a minimum monthly base fee: $500 at 200 or fewer peak concurrent users, $1,000 from 201 to 1,000, $2,000 from 1,001 to 2,000. A pilot burning 50,000 HD participant-minutes would owe $199.50 in media — and $500 in reality. That is 2.5 times the number in the business case, and volume discounts never touch it.

Stock Agora SDK vs custom Agora development

Most articles about Agora collapse the calling SDK and the platform around it into one thing. They are not one thing. The calling SDK is a commodity. What decides whether your product is defensible is everything wrapped around the call: token security, recording, AI agents, white-label theming, retention, moderation, observability and compliance.

The split is stable across every Agora project we have scoped. Agora sells you an edge network and a codec pipeline. You build the product.

Eight scenarios that force a custom Agora build

Multi-tenant white-label, recording with a retention policy, production AI voice agents, captions plus translation, moderation with an audit trail, more than 17 video hosts, hybrid CDN fallback, and per-tenant cost attribution — if two or more are on your roadmap you are already past the stock SDK. These are the moments teams actually call us, in rough order of how often they come up.

1. Multi-tenant white-label. One app, many customer brands, isolated channels, per-tenant theming and entitlements. The SDK has no tenant concept, so that layer is yours.

2. Recording with a policy. Ninety-day retention for one customer, seven years for a healthcare customer, automatic redaction of identifiers in transcripts, per-region buckets. None of that lives in the recording API.

3. AI voice agents in production. Routing across model vendors, conversation context that survives a dropped session, a fallback path when a provider degrades, and a meter that tells you the $0.10 a minute is earning its keep.

4. Real-time captions plus translation. Synchronised captions, speaker diarisation and a second translation pass. Translation is billed per target language, so the architecture decision and the cost decision are the same decision.

5. Moderation with an audit trail. Frame sampling on the live leg plus a slower, more expensive re-scan of completed recordings, a policy engine, and a log your trust-and-safety team can defend.

6. Past 17 video hosts. The opt-in 128-host tier changes your publish and subscribe roles, your audience-to-host promotion path, and your cross-channel relay design.

7. Hybrid CDN fallback. Above roughly 100,000 concurrent viewers you stop pushing everything through the real-time network and mirror to a CDN. Media Push starts at $1.99 per 1,000 minutes for a single host with no transcoding and climbs steeply once you mix hosts or change properties; the orchestration is yours either way. Our guide to scaling real-time video to a million viewers walks that boundary in detail.

8. A cost dashboard your CFO trusts. Cost per active user-minute by tenant, aggregate-resolution drift, recorded-resolution tiers, agent-minute burn. The Agora console shows usage; it does not show margin, and it cannot break usage down by individual user at all.

Agora’s technical envelope in 2026

An Agora channel supports up to one million concurrent users, but Agora recommends no more than 32 concurrent audio senders of which 17 also send video; a 128-host tier exists on request. Codecs run H.264 through AV1, end-to-end encryption is a restricted beta, and the platform carries ISO 27001 and a SOC 2 Type II claim. Here are those numbers with the caveats that matter.

Channel capacity. A channel supports up to one million concurrent users. The famous 17-host figure is narrower than people think: Agora’s documentation recommends no more than 32 concurrent audio senders per channel, of which up to 17 also send video. It is a recommendation, not an enforced cap — exceed it and streams start dropping for subscribers rather than throwing an error, which is a far nastier failure mode — Agora publishes no API for capping concurrent senders, so you enforce it yourself. The 128-host tier is opt-in: you have to ask support to switch it on, and inside it one subscriber can receive at most 50 remote hosts at once.

Latency. Agora publishes two figures it has never reconciled: “400 ms or less” worldwide on its current platform page, and a median under 76 ms globally in an undated network whitepaper. Both are Agora’s own, they differ by more than 5×, and neither is scoped to a workload. We plan against the pessimistic one and measure the real one ourselves.

Codecs. H.264, H.265, VP8, VP9 and AV1. AV1 and VP9 are still flagged beta in the Web SDK codec enum, and on native platforms AV1 is a support-gated extension rather than a default — the video release notes carry no AV1 entries at all. The widely quoted “AV1 is 42% smaller than H.264 and 25% faster than x264” numbers come from an AOMedia case page quoting an Agora engineer rather than an Agora benchmark, and Agora’s own Web release notes claim a more modest 30%. Keep H.264 and VP8 in the fallback list; forcing AV1 on a five-year-old Android fails to connect.

Encryption. Transport encryption is on by default. Media encryption is not — Agora’s own security page lists it as not applied by default, and you have to call the encryption API to switch it on, with aes-128-gcm2 or aes-256-gcm2 as the recommended modes. That distinction has bitten more than one compliance review. True end-to-end encryption is a third thing again, and it is still labelled beta, with restrictions most teams discover late: VP8 video and Opus audio only, Web client to Web client only with no cross-platform support, no bypass to server-side services, SEI metadata disabled, and it has to be switched on by Agora support. So E2EE rules out iOS and Android, and it rules out recording. If your compliance story needs both an encrypted call and a retained archive, that conflict gets resolved in architecture — usually with on-premise recording — not in a sales call.

Compliance. Agora publishes ISO/IEC 27001:2022, 27017 and 27018 certificates through DNV, plus a 27701, and its security white paper states SOC 2 Type II. Two things a diligence process should still do: check the certificate expiry dates, which currently run to 19 October 2026, and get the SOC 2 report itself under NDA rather than relying on the claim. On HIPAA, Agora offers compliance under a BAA as an option for healthcare customers but publishes no template and no on-request commitment, so treat it as a contract negotiation with a lead time, not a checkbox.

A reference architecture for a custom Agora build

Production Agora deployments converge on four tiers. Agora owns exactly one of them. Everything that makes the product defensible — and everything an auditor asks about — lives in the two below it.

Figure 3. The reference architecture we draw on every Agora scoping call. Agora owns the second row; rows three and four are the engagement.

Token service

The most-attacked surface in any Agora deployment. Tokens are short-lived, scoped to one channel and one role, and issued server-side after your own authentication. The App ID and certificate never ship to the client. We see 30-day unscoped tokens more often than we would like; a leaked one owns every channel for a month.

Recording orchestrator

Decides what gets recorded, in which mode, where the file lands, how long it stays and who can delete it. Cloud Recording at $5.99 per 1,000 HD minutes is the fast path; on-premise recording at $3.99 keeps decrypted media inside your own VPC and is both cheaper and easier to defend in an audit. That is an unusual combination, and it is why regulated products should look at on-premise mode first rather than last.

Moderation worker

Two pipelines, not one. The live leg samples frames at one to two per second and adds roughly 500 ms of latency. The async leg re-scans finished recordings with a heavier model and produces the record your trust-and-safety team acts on.

AI orchestrator

The bridge between Agora’s Conversational AI Engine and your product logic: prompt graph, conversation memory across sessions, provider fallback, and a meter. At $0.10 a minute the meter is not optional.

Conversational AI Engine: what $0.10 a minute buys

The Conversational AI Engine is Agora’s managed service for putting an AI voice agent into a call: it wires speech recognition, a language model and speech synthesis into an Agora channel as another participant. It costs $0.10 per agent-minute, with the first 300 minutes free. The number that circulated through 2025 — $0.0265 all-in — is no longer the price, and any cost model built on it understates voice-agent spend by roughly a factor of four.

What the $0.10 includes is the interesting part, and it has a trap in it. Agora’s pricing documentation is explicit that the unit price covers the selected ASR, LLM and TTS models, and you are charged the same price even if you bring your own key. The trap is the word “selected”: only Agora-managed models are inside the bundle. Pick a provider outside that list — ElevenLabs, for instance — and you pay $0.10 plus that vendor’s bill. Either way a negotiated contract of your own earns you nothing here.

The vendor coverage is genuinely broad. As of the June 2026 docs: ten ASR providers, nine LLM providers plus any OpenAI-compatible endpoint, and fourteen TTS providers. Speech-to-speech runs through OpenAI Realtime, Gemini Live (direct or via Vertex AI) or xAI Grok. Avatars ship through Akool, LiveAvatar, Anam and a generic connector, all still marked beta. Telephony — outbound calling — arrived in v2.0 on 15 November 2025 and is also beta and currently free, which makes its price a future surprise rather than a current one. MCP server support landed in v2.4 on 2 February 2026, letting the agent’s model call your tools directly. Read those vendor counts as integrations, not as the bundle: the list of models Agora actually manages keys for is a single-digit subset of each.

One honest caveat worth knowing before you build a roadmap on it: the engine entered private beta in February 2025 and public beta two weeks later, and Agora has never announced general availability. It has simply kept shipping versions — v2.5 on 31 March 2026, v2.8 on 11 June 2026, v2.9 on 1 July 2026. That is not a reason to avoid it. It is a reason to keep a provider abstraction in your orchestrator.

When to use the engine and when to build the chain yourself

Use the engine when you want a working voice agent in weeks and $0.10 a minute clears your unit economics. Build your own speech-to-text, model and speech-synthesis chain when you need a specific model, a cloned voice, strict data residency on the model provider, or when your volume is large enough that a negotiated LLM contract beats a flat bundle. The crossover is not a minute count — it is the point where your own negotiated rates land materially below $0.10 and you have engineers to keep the chain alive. Our multimodal agents on LiveKit playbook covers the same pattern on the open-source side, and our AI integration practice exists for exactly this seam.

Recording, moderation and compliance pipelines

Recording is the biggest source of both cost surprises and compliance failures on Agora projects. Three decisions drive everything downstream.

Composite versus individual. This is a storage and redaction decision, not a price decision — Agora charges the same recording rate either way. Composite mixes server-side into one file per session: cheapest to store, impossible to redact one participant afterwards. Individual writes one file per host: more storage, but a deletion request becomes a file delete instead of a re-encode. Multi-tenant platforms with legal-discovery obligations should take individual mode and stop arguing about it.

Cloud versus on-premise. On-premise recording is both cheaper ($3.99 versus $5.99 per 1,000 HD minutes) and the only mode where decrypted media never leaves your infrastructure. For anything touching health records this is the default, not the upgrade. Webpage recording, which captures a rendered browser tab, is $14 per 1,000 HD minutes — 2.3 times cloud recording and 3.5 times on-premise, and worth it only when the composition itself is the product.

Where the transcript lives. Real-Time Speech to Text at $16.99 per 1,000 minutes is expensive enough that transcript scope is an architecture decision. Transcribe the clinical part of a call rather than the waiting room, and write the result into your own store rather than reading it back out of a vendor.

Agora vs LiveKit, Daily, Twilio, Zoom and Chime

Every price below comes from the vendor’s own pricing page, fetched on 18 July 2026. Note row four especially: Twilio Programmable Video is alive. Twilio reversed the end-of-life on 21 October 2024 and still sells and invests in the product. Comparison articles that tell you to migrate off it are two years out of date.

All rates are normalised to dollars per participant-minute so the columns are actually comparable — Agora quotes per 1,000 minutes, everyone else quotes per minute.

Four things that table hides. Agora’s $0.00399 is the two-person case — the rate climbs with room size, which no competitor here does. LiveKit meters at least three things at once (participant-minutes, downstream bandwidth, and agent-session minutes at $0.01), so a bandwidth-heavy or agent-heavy product costs more there than the headline suggests; our LiveKit vs Agora cost analysis runs that math properly. Chime has no free tier at all, which quietly rules it out for pilots. And Cloudflare now owns Dyte, whose SDKs are in maintenance mode, so anything running on Dyte has a migration in its future. Zoom moved Video SDK onto a Build Platform credit model during 2025–2026, so check the current credit conversion before you model it.

Tsahi Levent-Levi, who has tracked this market independently for two decades at BlogGeek.me, put the industry-standard video API price at roughly $0.004 a minute in June 2026 and set a floor rather than a crossover: below about 1,500 to 2,000 average concurrent users, you are likely paying more on the build path, and his own conclusion is that building is not cheaper in the first year. That matches what we see. If you are weighing that specific question, our WebRTC vs Agora architecture tradeoffs and self-hosted Agora alternatives guide both go deeper than this section can.

Worked cost math: 40,000 telehealth consults a month

A telehealth platform running 40,000 consultations a month at 18 minutes each, two participants on Video HD, with composite recording for compliance and clinical notes transcribed on about 30% of minutes, costs $20,677 a month on Agora at 2026 list rates — $0.52 per consultation. Here is the arithmetic, because abstract pricing tables do not survive contact with a finance review.

Figure 4. From sessions to Standard minutes to a monthly invoice, with every multiplication shown. The AI agent line is what changes the answer.

Step one converts sessions into Standard minutes. 40,000 consults × 18 minutes × 2 participants = 1,440,000 participant-minutes. Each participant subscribes to exactly one 1280×720 stream, which is 921,600 pixels and lands in the HD tier, so the conversion ratio is 1:4: 5,760,000 Standard minutes.

Step two prices each line. RTC at the postpaid rate: 5,760,000 ÷ 1,000 × $0.99 = $5,702. Cloud Recording HD on the composite stream: 720,000 session-minutes ÷ 1,000 × $5.99 = $4,313. Speech to text is billed per unmuted host, not per session, so transcribing 30% of a two-person consult is 2 × 216,000 = 432,000 minutes ÷ 1,000 × $16.99 = $7,340. Analytics at the Standard tier: $449 flat.

Step three: $5,702 + $4,313 + $7,340 + $449 = $17,804 a month, or $0.45 per consultation.

Then add Cloud Proxy, because the clinicians sit behind hospital firewalls. Cloud Proxy is metered separately at the same resolution-tiered rates, with a monthly floor of $500 to $2,000 depending on peak concurrency — and the fee is whichever of the two is higher. Proxying just the clinician leg is 720,000 ÷ 1,000 × $3.99 = $2,873, well above the floor, so the metered number is what lands. Total: $20,677 a month, or $0.52 per consultation. The floor only bites small deployments, where a pilot burning 50,000 HD participant-minutes would owe $199.50 in proxied media and $500 in reality.

For a telehealth business charging $40 a visit, $0.52 is a little over 1% of revenue and the platform decision is not close. But note how far it moved. The model most teams build — 720,000 session-minutes at the headline HD rate, plus naive recording and analytics — lands around $0.19 a consult. Per-participant billing, per-unmuted-host transcription and separately metered proxying took it to $0.52. Closing that 2.7× gap is what this article exists for.

Now change one thing that sounds harmless. Add a nurse to the room. Every participant now subscribes to two streams, the aggregate hits 1,843,200 pixels, and the whole session bills at Full HD with a 1:9 ratio: 40,000 × 18 × 3 = 2,160,000 participant-minutes × 9 = 19,440,000 Standard minutes = $19,246. One extra person took the RTC line from $5,702 to $19,246. Nothing about the video changed.

Change one more. Put a $0.10-a-minute AI intake agent on every visit and you add 720,000 agent-minutes × $0.10 = $72,000 a month — three and a half times the entire rest of the bill. Run the agent on 10% of visits as a triage step instead and it costs $7,200, which is defensible. Same product, same vendor, a 10× difference in gross margin, decided in a design meeting.

Mini case: an Agora classroom shipped in one week

Situation. An EdTech customer needed a live classroom inside a product that already had its own courses, billing and identity. Building the classroom from raw WebRTC would have added months to a roadmap that had none, and the team had no real-time engineers.

Plan. We built the classroom as an embeddable component on Agora Flexible Classroom rather than on the raw SDK: video, interactive whiteboard, document editing across PDF, Word, Excel and PowerPoint, handwriting and notes, screen share, collaborative playback, and teacher-versus-student permission roles. The work that took the time was not the calling — it was the permission model, the theming, and wiring the component into an existing identity system so a student never sees a second login.

Outcome. The classroom went into the host product in about a week instead of a quarter, and Agora’s compliance posture came with it rather than being rebuilt. The trade-off is real and worth naming: buying Flexible Classroom means the media path, the billing tier and the feature roadmap belong to Agora. Our BrainCert engagement went the other way for exactly that reason — that platform runs its own WebRTC stack and now serves 100,000+ customers and 500 million+ classroom minutes at 99.995% uptime, with SOC 2 Type I and II and ISO 27001, none of which would have been possible on a component we did not control. Both calls were right for their product, and both were made in week one. Want that call made properly for your product?

What custom Agora development costs and how long it takes

A production-grade first cut on Agora runs eight to twelve weeks of senior engineering; a multi-tenant white-label build is twelve to sixteen, and a regulated platform sixteen to twenty plus hardening. At a $75 blended hourly rate, two engineers for eight to twelve weeks works out at 2 × 40 × 8 × $75 = $48,000 at the low end and $72,000 at the high end; the regulated shape roughly doubles that. Those are envelopes, not quotes — here is what moves them.

One thing genuinely compresses this: starting from a known-good reference architecture instead of discovering the token and recording design in week six. That is where the rework lives, and it is why we publish the architecture above rather than treating it as proprietary. If you want the counterpart view on staffing, our guide to hiring a WebRTC development company versus building in-house covers the trade-off in more depth.

The honest caveat: any number given before we have seen your average room size, your recording policy and your compliance obligations is decoration. Those three inputs move the estimate more than the platform choice does.

A decision framework in five questions

Five questions settle it: will any channel carry more than 17 video hosts, do you record under a retention or redaction policy, are your users behind firewalls, are AI voice agents on the roadmap, and where do your users physically sit. Three or more “yes” answers and you are budgeting a build, not an integration. We ask them in this order because it settles the answer fastest.

Figure 5. The five questions, in the order that resolves a scoping call fastest.

Q1. Will any channel ever carry more than 17 video hosts? If yes, you are in Interactive Live Streaming territory: the opt-in 128-host tier, custom publish and subscribe roles, and an audience-to-host promotion path. This is the single most common reason teams outgrow the stock SDK, usually around month four when someone sells a 30-speaker panel.

Q2. Do you record, and does anyone govern retention or redaction? If yes, you need an orchestrator with per-tenant policy, and probably on-premise mode. The recording API by itself has no opinion about your regulator.

Q3. Are your users behind corporate or hospital firewalls? If yes, Cloud Proxy goes on and becomes its own meter at the same resolution tiers as the call, with a $500 to $2,000 monthly floor underneath it. Small deployments pay the floor; anything at scale pays the meter, which in our worked example runs nearly six times the floor. Cost it now, not after the pilot.

Q4. Are AI voice agents on the roadmap in the next two quarters? If yes, price $0.10 a minute against your unit economics before anyone designs the feature. It is the one line item that can exceed everything else combined.

Q5. Where do your users actually sit? Heavy Southeast Asia, Latin America or Africa and Agora’s edge presence wins on join time and packet loss. Heavy US and EU with an agent-first product and you should price LiveKit and Daily alongside it. Already on Twilio and happy? Stay — there is nothing to migrate off.

Five pitfalls we see most often on Agora projects

1. Token sprawl. Thirty-day tokens, no role scoping, no channel binding. One leak and an attacker owns every channel for a month. Fix: five to twenty-four hour expiry, publisher or subscriber role, bound to one channel, mapped to your own user ID.

2. Modelling cost on published resolution. Nearly every spreadsheet we are handed multiplies session minutes by the encoder profile the team picked. Agora bills on subscribed aggregate, so the spreadsheet is right for two-person rooms and wrong for everything else. Fix: model per role and per room size, and re-model the day someone adds screen share.

3. Screen share left at the platform default. 1920×1080 on Windows, macOS and Web, pushed into every subscriber’s aggregate. Fix: set the capture resolution explicitly for the use case — a shared slide deck does not need 1080p — and check what subscribers actually receive rather than what you configured.

4. AV1 forced everywhere. It is still beta on Web and a support-gated extension on native. Forced onto an older Android the call does not connect. Fix: keep H.264 and VP8 in the codec list and let the SDK negotiate.

5. No network floor. A default HD profile on a weak mobile connection produces freezes and a churn spike. Fix: enable network quality control, fall back to audio-only below about 100 kbps, monitor packet loss through the SDK, and show the user what is happening.

KPIs that matter: quality, business, reliability

Quality KPIs. P95 join latency under 2.5 seconds. P95 packet loss under 2%. Audio mean opinion score above 4.0. Caption latency under 800 ms when transcription is on. These decide whether a user is still in the call at minute three.

Business KPIs. Cost per active user-minute, broken out by tenant. Thirty-day retention. Agent containment rate, meaning the share of AI-handled calls that never escalate to a human — at $0.10 a minute this is the number that decides whether the agent pays for itself. Recording storage as a share of total infrastructure spend.

Reliability KPIs. Session completion above 99.95%. Recording success above 99.5%. Detection of a platform incident inside 60 seconds through your own quality pipeline, not the vendor console. When Agora has a regional problem, you want to know before your customer tells you.

When NOT to build on Agora

Do not build on Agora when AI agents dominate the workload and you need the source code (pick LiveKit), when shipping in two weeks beats per-minute price (Daily), when you already run a working Twilio deployment, when your stack is AWS and the workload is voice-first (Chime), or when you have peer-to-peer pairs and the expertise to run TURN yourself. Here is the reasoning behind each.

Pick LiveKit when AI agents dominate and you want the code. The media server and the agents framework are both open source and self-hostable, which matters if you need to run in a customer’s VPC or negotiate your own model contracts instead of paying a bundled rate.

Pick Daily when shipping in two weeks beats per-minute price. Prebuilt UI, transparent tiering from $0.0015 a participant-minute at volume, and Pipecat for the agent layer.

Stay on Twilio if you are already there and it works. The end-of-life was reversed in October 2024, the product is actively invested in, and $0.004 a participant-minute for 50-person rooms is competitive. Migrating off a working platform to chase a rate card is how teams burn a quarter. If you genuinely do need to move, our Twilio Video migration guide compares the landing options honestly.

Pick AWS Chime when the stack is already AWS and the workload is voice-first. $0.0017 per attendee-minute for audio and video alike in a standard session is the simplest pricing in this comparison, and it does not escalate with room size the way Agora’s does. Two caveats: a high-definition session costs double and drops the participant cap from 250 to 25, and there is no free tier at all, so the pilot costs real money from minute one.

Pick raw WebRTC when you have peer-to-peer pairs and the expertise to run TURN. No per-minute bill, substantial engineering cost. Independent analysis puts the floor around 1,500 to 2,000 average concurrent users: below that, the payroll is more expensive than the invoice.

Is Agora a safe vendor to bet three years on

On the public numbers, yes, with caveats. Agora, Inc. trades on NASDAQ under API, so this is checkable rather than a matter of trust: first quarter 2026 revenue was $37.7 million, up 13.5% year over year, with 3,946 active customers as of 31 March 2026 and second-quarter guidance of $39 to $40 million.

Full-year 2025 revenue was $141.1 million and 2025 was the company’s first full year of GAAP profitability since 2018, with net income of $9.5 million against a $42.7 million loss the year before. Read that one carefully rather than celebrating it: the operating line was still negative in 2025 at about −$9.4 million, and the net profit came from roughly $15 million of interest earned on a cash pile of some $375 million at year end. A company earning its way to profitability through treasury income is stable, not compounding.

Three caveats a diligence process should catch. From Q1 2026 Agora stopped reporting its international and China segments separately, so year-over-year segment comparisons get harder from here. The $200 million buyback is the February 2022 authorisation running down — about 78% used — not fresh capital return. And on the open-source side, the TEN Framework that Agora promotes for multimodal agents is licensed under Apache 2.0 with additional conditions restricting on-device hosting and competitive redeployment; GitHub itself declines to classify the licence. It is source-available, not OSI open source, and it does not carry the exit guarantee LiveKit’s licensing does. None of this is disqualifying. All of it belongs in the risk register.

How to screen an Agora development partner

Screen an Agora partner on five things: whether they can explain aggregate-resolution billing, who owns their recording infrastructure, how they resolve the E2EE-versus-recording conflict, which projects they have moved off Agora, and whether they can show you a live cost meter. In that order, because it filters fastest. Most shops can wire a sample app; far fewer ship the layer above it.

1. Ask how Agora decides whether a call bills at HD or Full HD. The right answer is aggregate subscribed resolution — the sum of the streams each user receives — not the encoder profile you set. Anyone who has reconciled an Agora invoice knows this because it is the line that never matches the forecast. Anyone who has only read the pricing page will quote you $8.99 and stop.

2. Ask who owns the recording infrastructure. A partner who has only integrated the SDK has not solved retention, redaction, watermarking or audit evidence, and you will pay a second time to have it done.

3. Ask how they handle the E2EE-versus-recording conflict. Agora’s E2EE beta is Web-only, VP8 and Opus only, and cannot pass through server-side recording. A partner who has shipped regulated video will name those restrictions unprompted and describe the workaround. A partner who has not will tell you both are supported.

4. Ask which projects they moved off Agora. Migration scars are the experience that tells you which platform fits which workload. A team that has only ever shipped on one platform cannot give you a comparison, only a preference.

5. Ask to see a cost meter from a live product. Cost per active user-minute by tenant is the artefact that separates teams who ship products from teams who ship integrations. Ours is a standing part of our video conferencing practice, and the Learn: video streaming track walks through the underlying delivery decisions if you want to go deeper before you hire anyone.

Frequently asked questions

How much do Agora custom development services cost?

Roughly $48,000 to $72,000 for a production first cut: eight to twelve weeks of two senior engineers at a $75 blended hourly rate, covering a hardened token service, recording orchestrator, analytics and one AI feature. A stock SDK integration is two to three weeks of one engineer. A multi-tenant white-label build is twelve to sixteen weeks and a regulated platform sixteen to twenty plus hardening, so roughly double. Your average room size, recording policy and compliance obligations move the number more than the platform choice does.

What does Agora cost per minute in 2026?

Audio-only is $0.99 per 1,000 participant-minutes, Video HD is $3.99, Full HD is $8.99, 2K is $15.99 and 2K+ is $35.99. Cloud Recording HD adds $5.99 per 1,000 recorded minutes, speech to text is $16.99, and the Conversational AI Engine is $100 per 1,000 agent-minutes. Minutes are billed per participant and each account gets 10,000 free Standard minutes a month. The tier is set by the total resolution each user subscribes to, so the same 720p call bills at HD for two people and at 2K+ for six.

Is the Agora Conversational AI Engine really $0.10 a minute?

Yes. Agora's pricing documentation lists the Conversational AI Engine audio task at $0.10 per minute and states that the price includes the selected ASR, LLM and TTS models — you pay the same rate even if you bring your own API keys. The first 300 minutes are free. The $0.0265 figure that circulated in 2025 is no longer current.

Is Agora HIPAA compliant?

Not by default. Agora offers HIPAA compliance under a Business Associate Agreement as an option for healthcare customers, and publishes ISO/IEC 27001:2022, 27017, 27018 and 27701 certificates plus a SOC 2 Type II claim in its security white paper. There is no public BAA template or on-request commitment, so budget negotiation time. The compliant architecture is still yours to build: on-premise recording so decrypted media stays in your VPC, geo-fenced routing, customer-managed keys and a complete audit log on your side.

Did Twilio shut down Programmable Video?

No. Twilio announced an end-of-life in March 2024, then reversed it on 21 October 2024, stating that Twilio Video will remain a standalone product it continues to invest in. It is sold today at $0.004 per participant-minute with a 50-participant Group Room limit. Any article telling you to migrate off Twilio Video because it is being sunset is out of date.

How many people can join an Agora channel?

A channel supports up to one million concurrent users. The constraint people hit is publishers, not viewers: Agora recommends no more than 32 concurrent audio senders per channel, of which up to 17 also send video. A 128-host tier exists but must be enabled by Agora support. One subscriber can receive at most 50 remote hosts simultaneously.

Can I replace Agora with plain WebRTC?

Yes, and Agora is itself built on WebRTC — so the real choice is a managed edge network versus running your own TURN and SFU infrastructure. Raw WebRTC removes the per-minute bill and adds substantial engineering cost. Independent analysis sets a floor rather than a crossover: below roughly 1,500 to 2,000 average concurrent users you are likely paying more on the build path, and building is rarely cheaper in the first year.

Does Agora support end-to-end encryption?

Yes, but only as a restricted beta: VP8 and Opus only, Web client to Web client only with no iOS or Android support, no server-side services such as recording or transcoding, SEI metadata disabled, and it has to be switched on by Agora support. Separately, transport encryption is on by default while media encryption is not — that one has to be enabled through the SDK using aes-128-gcm2 or aes-256-gcm2. If your compliance story needs both an encrypted call and a retained archive, resolve it in architecture, usually with on-premise recording.

What is Agora.io?

Agora, Inc. (NASDAQ: API) is a real-time engagement platform that sells SDKs for video calling, voice, interactive live streaming, signaling, chat and AI voice agents, delivered over its own private edge network called SD-RTN. It reported $141.1 million of revenue in 2025 and 3,946 active customers as of 31 March 2026. It is not related to Agora Financial, the Agora blockchain project, or the ancient Greek marketplace of the same name.

Does Agora have a free tier?

Yes. Every Agora account gets 10,000 free Standard minutes a month across Video Calling, Voice Calling, Interactive Live Streaming and Broadcast Streaming, and the Conversational AI Engine carries a separate 300 free minutes. One catch: those 10,000 free minutes stop applying once you buy a prepaid package.

Is Agora cheaper than LiveKit?

For plain two-person video, usually yes: Agora is about $0.00399 per participant-minute at HD against LiveKit Cloud's $0.0005 per participant-minute plus $0.12 per GB of downstream bandwidth, and bandwidth is where LiveKit bills add up. The order flips for AI-agent workloads, where LiveKit charges $0.01 per agent-minute against Agora's $0.10, and for large rooms, where Agora's aggregate-resolution tiers escalate and LiveKit's rate does not.

How long does an Agora MVP take to ship?

A stock SDK MVP with one-to-one calling and a basic UI ships in two to four weeks. Add a hardened token server, recording and a simple AI agent and it is eight to twelve. A multi-tenant, regulated, agent-enabled product is a twelve to twenty week first cut with another eight to twelve weeks of hardening before a serious launch.

What to read next

Ready to make Agora work for your product?

Custom Agora development services exist because the calling SDK is the easy part. The token service, the recording orchestrator, the moderation worker, the AI orchestrator and the per-tenant cost meter are what make a real-time product defensible under load, defensible under audit, and profitable in month six.

Price the four line items that actually move: aggregate subscribed resolution, transcription scope, recorded-resolution tiers, and separately metered Cloud Proxy. Price the AI agent before you design it, because $0.10 a minute can dwarf everything else. Then pick the platform on geography and workload rather than on a rate card — and if the honest answer for your product is LiveKit, Daily or staying on Twilio, we would rather say so on a call than find out together in month four.