The Models Behind Your AI

$10 / $50 per 1M tokens (input / output), 1M context, 128K max output. Thinking is always on; effort is tuned with output_config.effort rather than a token budget.

$5 / $25 per 1M tokens, 1M context, 128K max output. Adaptive thinking on by default.

Verified in Microsoft Learn catalog, version 2025-04-16. 200,000 input / 100,000 output. ~$1.10/1M input, $4.40/1M output; reasoning tokens bill as output. gpt-5-mini is cheaper on both input and output with a larger context window, so o4-mini is mostly for existing deployments.

Verified in the Microsoft Learn Foundry catalog and OpenAI's GPT-5.6 announcement. ~$2.50/1M input, $15/1M output. Same 1,050,000 context, same 272k long-context billing threshold and same tool-calling caveat as Sol. Also has built-in web search, file search and computer use — but OpenAI's hosted web search bills $10 per 1k calls PLUS retrieved-content tokens, so budget that separately. Good pattern: run Terra, escalate to Sol only when it stalls.

Verified in Microsoft Learn catalog, version 2026-03-05. 1,050,000 context / 128k output, training data to August 2025. ~$2.50/1M input, $15/1M output. Batch API eligible, which cuts cost further for jobs that don't need an instant answer.

Verified in Microsoft Learn catalog, version 2025-08-07. 400,000 context / 128k output, training data to 2024-09-30. ~$1.25/1M input, $10/1M output. REGISTRATION IS REQUIRED for gpt-5 (aka.ms/oai/gpt5access); the mini/nano variants are not gated. Available for Provisioned Throughput Units.

Verified in Microsoft Learn catalog, version 2025-04-16. 200,000 input / 100,000 output. ~$2/1M input, $8/1M output. Reasoning tokens bill as output, so real cost runs above the headline rate. OpenAI no longer ships frontier capability in the o-series — prefer GPT-5.4/5.6 for new work.

Verified in Microsoft Learn 'Foundry Models sold by Azure' (page updated 2026-08-27) and on OpenAI's own launch page. ~$5/1M input, $30/1M output. 1,050,000 context / 128k output. Billing gotcha: prompts over 272k input tokens reprice the WHOLE request at long-context rates. Chat Completions supports function tools only when reasoning_effort is 'none' — use the Responses API for tool calling. Some quota tiers require a quota request.

Verified in Microsoft Learn catalog, version 2026-04-24. ~$5/1M input, $30/1M output. DOCUMENTED GOTCHA: on the Responses API the effective combined prompt+generation budget is ~922,000 tokens, lower than the advertised 1,050,000 window — you cannot pair a 922k prompt with 128k output. Overruns return HTTP 200 with status 'incomplete' rather than an error, so check for it explicitly.

Model id and 2026-03-05 version verified in the Microsoft Learn catalog. Responses API ONLY — no Chat Completions. COST TIER IS AN ESTIMATE: no verified per-token price published for gpt-5.4-pro; tier 4 is inferred from the o3-pro / GPT-5-Pro pricing pattern. Confirm on the Azure pricing page before budgeting.

$1 / $5 per 1M tokens, 200K context.

$2 / $10 per 1M tokens, 1M context. Strong quality-per-dollar for production traffic.

Verified in Microsoft Learn catalog, version 2026-03-17. 400,000 context / 128k output. ~$0.20/1M input, $1.25/1M output. Unlike its siblings it does NOT list computer use in its capability set.

Verified in Microsoft Learn catalog, version 2025-08-07. 400,000 context / 128k output. ~$0.05/1M input, $0.40/1M output — confirmed by multiple sources as the cheapest Azure OpenAI text model. No registration required.

Verified in Microsoft Learn catalog, version 2025-04-14. ~$0.10/1M input, $0.40/1M output — second-cheapest Azure text model after gpt-5-nano. Same 300k-standard-deployment ceiling and large-tool-definition known issue as the rest of the 4.1 series.

Verified in Microsoft Learn catalog, version 2024-07-18. 128k input / 16,384 output. ~$0.15/1M input, $0.60/1M output. LEGACY — originally the GPT-3.5 Turbo replacement. gpt-5-nano is cheaper with a 400k context window.

Model id verified in Microsoft Learn catalog, version 2026-07-09, with the same 1,050,000 context window as Sol/Terra. PRICING DISPUTE: OpenAI's launch coverage lists $1/1M in and $6/1M out (tier 2), while OpenAI's pricing docs and one roundup list ~$0.20/$1.20 (tier 1). I set tier 2 as the conservative read — confirm on the Azure pricing page for your region and deployment type before budgeting.

Verified in Microsoft Learn catalog, version 2026-03-17. 400,000 context / 128k output. ~$0.75/1M input, $4.50/1M output. Batch eligible. Limited to specific deployment types (Global Standard or provisioned) — check the region availability table before assuming Standard regional works. Foundry Agent Service support has lagged general availability.

Verified in Microsoft Learn catalog, version 2025-08-07. 400,000 context / 128k output. ~$0.25/1M input, $2/1M output. No registration gate, unlike full gpt-5. Broadest region availability of the reasoning-capable tier.

Verified in Microsoft Learn catalog, version 2025-04-14. ~$0.40/1M input, $1.60/1M output. Same 300k-on-standard context ceiling and the same large-tool-definition known issue as gpt-4.1. Compare against gpt-5-mini, which costs less on input and adds reasoning.

Verified in Microsoft Learn catalog, version 2025-04-14. ~$2/1M input, $8/1M output. Context is nominally 1,047,576 but PRACTICALLY 300,000 on standard deployments and 128,000 on provisioned-managed and batch. DOCUMENTED KNOWN ISSUE across the 4.1 series: tool/function definitions over 300k tokens fail with context_length_exceeded (Chat Completions) or a bare HTTP 500 (Responses API). Provisioned deployments need spillover configured for long-context requests.

Verified in Microsoft Learn catalog; versions 2024-11-20, 2024-08-06, 2024-05-13. 128k input / 16,384 output, training data to October 2023. ~$2.50/1M input, $10/1M output. LEGACY — gpt-5 costs half as much on input and thinks better. Check the model retirement schedule before building on it.

Verified on ai.google.dev and blog.google. Released ~2026-08-13; Google's most capable Flash model, built for agentic workflows. $0.75/$3.75 per 1M is INTRODUCTORY through 2026-12-31 and DOUBLES to $1.50/$7.50 on 2027-01-01 — model your run-rate on the 2027 price. Context caching $0.075/1M plus $0.50/1M/hour storage. Batch API 50% off. NOT in Azure AI Foundry.

Present in the Microsoft Learn Foundry model list, with an important restriction: it is available ONLY through Foundry Agent Service via the Deep Research tool — you cannot create a normal model deployment and call it from the standard chat/responses endpoints. No capability table or per-token price is published alongside the other models, so the cost tier is an estimate based on deep-research runs consuming very large token volumes. Verify regional availability before designing around it.

SaaS, not on Foundry. Returns short LLM-ready context rather than full pages — the wrong tool if you need complete document text. 1,000 free searches/month; ~$30/mo for 4,000 credits or $0.008/credit pay-as-you-go, dropping to ~$0.005/credit on the ~$500/mo Growth plan. Pricing is from third-party comparisons, one of which is authored by a competitor.

Verified directly from jina.ai/reader. Handles PDFs natively, captions images with a vision model, supports CSS-selector targeting/exclusion, and s.jina.ai does search-plus-content. Free without a key at 20 RPM; free key gives 500 RPM plus 10M complimentary tokens; paid is per output token (widely cited ~$0.02 per 1M). Known limitation: lightweight rendering can return incomplete content on JavaScript-heavy sites — use Firecrawl for those.

A SaaS service you call from your own code, not a model you deploy. Uses full Playwright rendering, which is the main reason to pick it over Jina on JavaScript-heavy sites. Pricing from third-party reviews: Free 1k credits, Hobby ~$16/mo, Standard ~$83/mo for 100k credits (~$0.00083/page), Growth ~$333/mo, Scale ~$599/mo. Repeatedly flagged watch-outs: credits do not roll over, Stealth Mode costs 5x per page, and the FIRE-1 agent bills even on failed runs. Its published 94.7% SimpleQA figure is a vendor self-benchmark.

A SaaS search API, not a Foundry model. Embeddings-based search over its own index, which is the differentiator versus keyword/SERP wrappers. Roughly $7 per 1k search requests; $10 free credits monthly on the free tier, $20 signup credit, Websets plan ~$49/mo for 8k credits. Pricing from vendor summaries and third-party reviews rather than a direct docs fetch.

COST TIER CORRECTED this pass: Brave REMOVED its free tier in February 2026. Search is now ~$5 per 1,000 requests (50 RPS); the Answers plan is ~$4 per 1k web searches plus ~$5 per 1M tokens. All plans include $5/month in credits (~1,000 queries) — that is a credit allowance, not a free plan, and a saved card is billed past it. Returns metadata and snippets rather than full cleaned page content, so pair it with a reader like Jina or Firecrawl if you need document text.

Verified on ai.google.dev, cloud.google.com and the DeepMind model card. Released 2026-02-19 and still Google's current Pro model. ~$2/$12 per 1M under 200K tokens, $4/$18 over; 1M context. Search grounding gives 5,000 free prompts/month across Gemini 3.x, then $14 per 1k queries. IMPORTANT: Gemini 3.5 Pro has missed three announced dates (late June, 17 July, early August) and is still unreleased as of late August 2026 — do not plan around it. NOT in Azure AI Foundry.

SaaS, not on Foundry. ~$3/$15 per 1M tokens PLUS a per-request search fee of roughly $6-$14 per 1k requests scaled by search_context_size — the request fee, not the tokens, is usually the real bill. The cheaper 'sonar' tier is ~$1/$1 per 1M plus ~$5-$12 per 1k requests. Pricing from third-party trackers and community posts; verify against docs.perplexity.ai before committing.

From xAI. Model id 'grok-4.6' confirmed present in the Microsoft Learn 'Foundry Models sold by Azure' list (page updated 2026-08-26), alongside grok-4, grok-4.1 and grok-4.3. Landed in Foundry Models in PUBLIC PREVIEW on 2026-08-26 at $2 input / $6 output / $0.50 cached per 1M, matching xAI's direct pricing. 500K context (not 1M); over 200K tokens the whole request reprices to $4/$12. Microsoft's own Foundry blog calls out data extraction as a target use case. Preview status means no SLA.

Verified on OpenAI's pricing page and in the Microsoft Learn Foundry image-generation table. $2.00/M text input, $2.50/M image input, $8.00/M output — roughly a quarter the output cost of gpt-image-2. Commonly cited around $0.005-$0.05 per image depending on size and quality. Does NOT support image edits or input fidelity — generation only.

Stability AI, built on an enhanced SDXL rather than SD 3.5. Same caveat as Stable Image Ultra: present in the Foundry catalog via the Stability partner registry rather than as a Microsoft-sold model, announced early 2025 and not re-verified as current for August 2026 — check the portal. Positioned for speed and efficiency over maximum fidelity.

Black Forest Labs. Verified in Microsoft Learn. Text input only (5,000 tokens), 1.6 MP max output. Available via the OpenAI-compatible Image API (images/generations) as well as the BFL provider API. Microsoft's description emphasizes fast inference, strong prompt adherence and competitive pricing. Default quota 6 RPM. Older generation than FLUX.2 — pick it for speed and cost, not for the quality ceiling.

Microsoft AI. Verified in the Microsoft Learn Foundry catalog and the ai.azure.com model catalog; Preview. Same 32k context, same ~1 MP output ceiling and 768x768 minimum as the Pro variant. Foundry also lists plain MAI-Image-2.5 and MAI-Image-2e in preview. Reported base rates around $5/M text input, $8/M image input, $47/M image output for the 2.5 family — not confirmed on an official Azure pricing page, so treat exact pricing as medium confidence while availability is high.

Vendor is Stability AI; powered by Stable Diffusion 3.5 Large. Confirmed present in the Azure AI Foundry catalog under the azureml-stabilityai registry, with an Azure Marketplace SaaS offer, billed pay-as-you-go per output image. Confidence stays medium on purpose: this is a PARTNER listing, not a model sold directly by Azure, so it does not appear in the Microsoft-managed model list I verified, and the announcements date from early 2025. Check the portal before committing. Stability's image line has lost ground to FLUX.2 in 2026 photorealism comparisons.

Verified on OpenAI's pricing page as the current flagship image model AND in the Microsoft Learn Foundry image-generation table (max request 4,000 characters). Released ~April 2026; reported to run a reasoning pass before generating. $5.00/M text input, $8.00/M image input, $30.00/M output tokens; batch 50% off. Real-world per-image commonly cited at ~$0.005-$0.21 depending on quality tier and resolution. Best Azure availability of any frontier image model here.

Verified on OpenAI pricing, in the Microsoft Learn Foundry image table, and in the December 2025 'what's new' entry. Text-to-image, image-to-image editing, inpainting, face preservation, up to 1024x1536 / 1536x1024. Microsoft's Foundry blog cites ~4x faster generation and ~20% lower cost than gpt-image-1. $5.00/M text input, $8.00/M image input, $32.00/M output. REQUIRES a limited-access application (aka.ms/oai/gptimage1.5access) — you cannot just deploy it.

From Black Forest Labs. Verified in Microsoft Learn 'Deploy and use FLUX models in Microsoft Foundry' (updated 2026-07-24): model ID FLUX.2-pro, provider path /providers/blackforestlabs/v1/flux-2-pro. Public Preview on Azure, global standard deployment in all regions. 4 MP max output, 32k token context, up to 8 reference images (multi-reference works via API but NOT in the Foundry playground). Rate limits 15/30/100 RPM by tier. IMPORTANT: Foundry provides NO built-in content filtering for FLUX at deployment time — you must wire up Azure AI Content Safety yourself.

Black Forest Labs. Verified in the same Microsoft Learn FLUX doc: model ID FLUX.2-flex, path flux-2-flex, Public Preview. 4 MP max, up to 10 reference images. Exposes guidance (1.5-10, default 4.5) and steps (max 50) that FLUX.2 [pro] does not surface the same way. Notably lower quota than pro: 5/10/25 RPM by tier — check that against your throughput needs. Same no-built-in-content-filtering caveat as FLUX.2 [pro].

Vendor is Microsoft AI (MAI). Verified in the Microsoft Learn 'Foundry Models sold by Azure' list (updated 2026-08-26), a dedicated 'Deploy and use MAI image models' how-to, and the ai.azure.com catalog. Preview. Text-to-image and image-to-image, 32k token context, PNG output. Strengths: identity and character consistency across restyling, localized edits that leave the rest untouched, and structured document/diagram/slide visuals. NOTABLE LIMIT: max total output is 1,048,576 pixels (~1024x1024), minimum 768x768 — it cannot match FLUX.2's 4 MP or Nano Banana's 4K.

CONFIRMED this pass (previously medium): announced by Google as the fastest and most cost-efficient model in the Nano Banana family, covered on the Google Cloud blog and by VentureBeat, which reports ~4-second enterprise generations. Available via Gemini API / AI Studio / Vertex. I did not verify an exact per-image rate, so the tier is positional — it sits below Nano Banana 2. NOT on Azure AI Foundry.

CONFIRMED this pass (previously medium). Black Forest Labs, released 2026-01-16 as the compact on-device member of the FLUX.2 family. Weights on Hugging Face (FLUX.2-klein-4B, FLUX.2-klein-9b-kv), official inference repo at github.com/black-forest-labs/flux2, and an NVIDIA NIM listing. THE 4B MODELS ARE APACHE 2.0 — the earlier 'check the licence before commercial use' warning does not apply to the 4B tier; verify separately for the larger variants. 4B fits in ~13GB VRAM (RTX 3090/4070 and above); BFL ships FP8 and NVFP4 builds, with sub-5s generation demonstrated on 8GB VRAM on ASUS ProArt RTX laptops. NOT in Azure Foundry — Foundry carries only FLUX.2-pro, FLUX.2-flex, FLUX.1-Kontext-pro and FLUX-1.1-pro. Cost tier 1 means you pay compute, not per image.

LICENSING CORRECTED this pass: Ideogram's own press release (ideogram.ai/news/ideogram-4.0/) describes it as an open-weight release WITH A COMMERCIAL LICENCE, downloadable today, with enterprises free to fine-tune on their own brand data and deploy in their own environment — the earlier 'non-commercial weights' warning was wrong, though you should still read the licence text yourself. Released 2026-06-03; 9.3B single-stream Diffusion Transformer, native 2K text rendering, reported 0.97 OCR accuracy, structured JSON prompts for layout/colour/text placement, runs on a 24GB GPU. Paid paths: Ideogram API ~$0.03-$0.10/image, subscription ~$7-$42/mo, or fal.ai ~$0.03-$0.10/megapixel. Not on Azure Foundry.

A royalty-free photo search source, NOT a generative model. Free: 50 requests/hour in Demo, 5,000/hour once approved for Production, still free and commercial use included. LICENSING SUBTLETY THAT CATCHES PEOPLE OUT: the Unsplash Licence itself does not require attribution, but the API TERMS do — every displayed image must credit Unsplash and the photographer with a clickable link to the photographer's profile, and failing to do so can get your API access revoked. Risk note: no model releases and no indemnification, so you carry the legal risk for recognizable people, trademarks and property.

A royalty-free search source, NOT a generative model. Free: 200 requests/hour and 20,000/month by default, and Pexels will lift the rate limits entirely at no cost if your platform shows acceptable attribution to Pexels and its contributors. The Pexels Licence permits free commercial use without attribution, but the API guidelines ask you to credit photographers ('Photo by X on Pexels' linking to the photo page) or at minimum 'Photos provided by Pexels'. Same risk profile as Unsplash: no model releases, no indemnification.

A royalty-free search source, NOT a generative model. Free: 100 requests per 60 seconds, keyed to your API key rather than IP. The differentiator: unlike Pexels and Unsplash, Pixabay does NOT require attribution for API usage. TWO HARD TECHNICAL CONSTRAINTS: responses must be cached for 24 hours, and permanent hotlinking is prohibited — webformatURL values expire after 24 hours, so you must download images into your own storage. Budget for a blob-storage copy step. Quality is more variable than Unsplash or Pexels, and the same no-model-release / no-indemnification risk applies.

Verified on blog.google, deepmind.google and Google Cloud docs. Preview released 2026-02-26, full release 2026-06-18 — this is NEWER than Nano Banana Pro, so do not assume 'Pro' means latest. 512px to 4K; per-image ~$0.045 (0.5K), $0.067 (1K), $0.101 (2K), $0.151 (4K); token-metered listings show $0.50/M input, $60/M image output. Batch is a flat 50% off. Adds web-search visual grounding, character consistency for ~5 characters, fidelity across ~14 objects. NOT on Azure AI Foundry.

CRITICAL INTEGRATION CAVEAT, corroborated across multiple sources including Midjourney's own docs: there is still NO official public REST API as of August 2026. Access is Discord or midjourney.com from ~$10/month; an Enterprise API has only been floated in a survey and nothing broadly available has shipped. Every 'Midjourney API' you find is an unofficial third-party wrapper that violates the terms of service and can break without warning — do not design an automated pipeline around it. Version history confirmed: v8.0 alpha 2026-03-17 (alpha site closed 2026-07-24), v8.1 2026-04-30 (2K HD), v8.2 default since 2026-07-24. Not on Azure Foundry.

'Nano Banana' is Google's own nickname for its Gemini image line, not a third-party product — this is Gemini 3 Pro Image, launched 2025-11-20, confirmed on deepmind.google and blog.google. ~$0.134/image at 1K-2K, ~$0.24 at 4K standard; batch/flex ~50% off. Best-in-class in-image text rendering, localized edits, camera and lighting control, blends up to 14 reference images and holds up to 5 people consistent. NOT on Azure AI Foundry — Gemini API / AI Studio / Vertex AI only. Effectively tier 4 if you generate mostly at 4K.

CONFLICTING LIFECYCLE INFORMATION — verify in the portal before committing. Still listed (flagged NEW, in preview, max request 4,000 characters) in Microsoft Learn's 'Foundry Models sold by Azure' page as updated 2026-08-27, AND has a live ai.azure.com catalog entry. But OpenAI shut the consumer Sora app on 2026-04-26, the public API has a scheduled sunset of 2026-09-24, and Microsoft Q&A threads show customers being told the Foundry deployment (v2025-12-08) retires earlier still — dates of 2026-06-02/06-06 and 2026-09-14 all appear in different Microsoft communications. The earlier 'sora' v2025-05-02 was already retired 2026-02-28. Variants sora-2 and sora-2-pro, v1 API at /openai/v1/video/generations/jobs, ~$0.10-$0.70 per second. There is no successor video model hosted in Foundry to migrate to. Do not start new work here.

CONFIRMED this pass. Google's 'any-to-any' generative media family announced at I/O 2026 (2026-05-19), opened to developers 2026-06-30 via the Gemini API and AI Studio, with the 1.1 Flash update landing 2026-08-28 (blog.google, ai.google.dev/gemini-api/docs/omni, DeepMind model card). 1.1 adds scene extension to 40 seconds, start/end keyframe control, a faster draft mode and 4K output — note Google's own docs say 1080p and 4K are UPSCALED, not generated natively. Base rate ~$0.10 per second of output. modelId left empty deliberately: the 1.0 API string was gemini-omni-flash-preview and I could not confirm the exact 1.1 identifier — check ai.google.dev before coding against it. NOT in Azure AI Foundry, and never will be.

From Kuaishou Technology (the schema has no Kuaishou option, hence vendor 'Other'). Launched 2026-02-04 as a rebuilt multimodal architecture. Specs: native 3840x2160 4K (not upscaled), up to 60fps, 15-second clips, multi-shot storyboarding with up to 6 cuts per generation, native audio with multilingual lip-sync. Consistently cheapest of the major models at ~$0.09-$0.14/sec. A speed-optimized 3.0 Turbo shipped 2026-06-17 alongside a higher-fidelity 3.0 Pro. Confidence is medium because every detail comes from secondary aggregator coverage rather than Kuaishou's own documentation. Chinese vendor — consider data-residency and procurement implications. Not in Azure Foundry.

CONFIRMED this pass. Previewed 2026-06-23 at Volcano Engine's FORCE conference and the API went live shortly after (CineD, Picsart, Morphic, OpenArt all carry it), so treat late-June-to-July 2026 as the release window. Renders a full 30-second clip in a single generation with no stitching or extension passes; accepts up to 50 reference inputs (30 images, 10 videos, 10 audio); native 1080p with 4K max; text-to-video, image-to-video and reference-to-video; native audio in 10+ languages; region-level editing. Access is via ByteDance's own consoles — BytePlus ModelArk (international) and Volcano Engine Ark (China). modelId left empty: ids like 'doubao-seedance-2.5' circulate in secondary sources but I could not confirm them against ByteDance's own docs. Billing is token-based (~¥70/M tokens without video input), NOT per-second; tier extrapolated from Seedance 2.0's ~$0.09/sec 480p and ~$0.20/sec 720p. The older Seedance 2.0 is still maintained and got a 4K update in June 2026 if you want the cheaper tier. Note: Seedance went viral generating clips of real actors and drew significant copyright criticism in China — a real consideration for commercial use. Not in Azure Foundry.

Verified against runwayml.com's own research announcement and help centre rather than only third-party blogs. Available via the Runway API during 2026 (sources disagree on the exact date — announced 2025-12-01, API availability variously reported 2026-02-10 and 2026-03-25). Strongest documented differentiator is adherence to long, sequenced, multi-clause instructions; it took the #1 spot on Artificial Analysis's Video Arena at launch. 1080p output with integrated audio generation, and it has since added native audio editing of existing videos. The older Gen-4 / Gen-4 Turbo tiers remain available and cheaper (up to 10s, up to 4K) if you don't need 4.5's prompt adherence. Runway also launched a developer platform, Runway Dev, in July 2026. Reported around $1.50 per clip. Not in Azure Foundry.

Still the current Veo — VEO 4 DOES NOT EXIST as of August 2026. Multiple SEO blogs speculate about a 'Veo 4 release date'; Google has published no Veo 4 model page, API model ID, model card or pricing. Treat any Veo 4 claim as unverified. Veo 3.1 shipped October 2025 with Fast and Lite variants below it. Reported ~$0.75/sec for Standard — the most expensive per-second rate among major models, hence tier 4; Fast is ~$0.15/sec. Widely cited as shipping native audio in the output rather than requiring a separate audio pass. Google Cloud only, not Azure Foundry.

CONFIRMED this pass: launched 2026-03-16 per D-ID's own newsroom and the PR Newswire release. A different product shape from HeyGen and Synthesia — streaming interactive agents rather than batch video rendering. Diffusion-based model trained on real actor performances, sub-0.5-second conversational turns, lip-sync at up to 4K, real-time streaming API reported at up to 100 FPS, pluggable custom knowledge sources and LLMs, and dynamic alignment of tone to the selected sentiment. Also retains D-ID's original strength of animating still photos. Cheapest entry point of the three avatar platforms — available on all D-ID plans from around $5.90-$5.99/month, with agent usage billed at 0.5 credit per 30 seconds of generated response. Not in Azure Foundry.

Still current and still documented in HeyGen's help centre — not deprecated by Avatar V, because it solves a different input problem (one photo versus a 15-second video clip). HeyGen's own framing is that Avatar IV output is 'recognizable' while Avatar V is 'indistinguishable', so this is the cheaper, lower-fidelity tier. Includes voice sync, facial expressions and hand gestures. The natural fallback path if you want to degrade gracefully rather than hard-fail when the newer engine is unavailable. Not in Azure Foundry.

Confirmed on HeyGen's own announcement blog and product page. Builds a studio-quality avatar from a ~15-second clip with strong identity consistency. The genuinely novel part is separating performance from appearance, which maps directly onto a requirement to vary outfits, looks and backgrounds per piece of content rather than reusing one look — that is native here rather than something to hack around. Shipped in HeyGen's May 2026 release wave alongside HyperFrames, LiveAvatar and Android support. If you already pass Avatar Look IDs as the HeyGen avatar_id for video creation, verify how Avatar V looks are addressed before switching. Not in Azure Foundry; HeyGen is a standalone SaaS/API.

CONFIRMED CURRENT this pass — I searched synthesia.io specifically for an 'Express-3' and found none, so do not assume a newer number exists. Express-2 launched September 2025 and is now available on all paid Synthesia plans. Diffusion transformer video model paired with an ~800M-parameter voice-cloning engine, producing 1080p 30fps full-body avatars with natural hand and body gestures rather than head-and-shoulders lip-sync, at arbitrary length. Ships within the broader Synthesia 3.0 platform (Video Agents, Courses, Copilot, AI Dubbing). Known weakness: diffusion-based skin rendering reads as hyper-smooth in tight framing, improving in medium and wide shots. Seat/enterprise licensing rather than per-second API billing, so the cost tier is a rough relative placement. Not in Azure Foundry.

Verified in the Microsoft Learn catalog and on OpenAI pricing ($0.13 per 1M). 8,192 max request tokens, 3,072 output dimensions. Best MIRACL (54.9) and MTEB (64.6) of the OpenAI embedding models, though it is a 2024-generation model now outranked by newer entrants — you trade a few points of retrieval quality for first-party Azure deployment and governance. Supports the 'dimensions' parameter to shrink vectors. You CANNOT upgrade between embedding models: switching means regenerating every embedding. Max 2,048 array items per call.

Verified in the Microsoft Learn catalog and on OpenAI pricing ($0.02 per 1M) — the cheapest model of any kind in this list. 8,192 max request tokens, 1,536 dimensions. MIRACL 44.0 / MTEB 62.3, about 6.5x cheaper than 3-large for a modest quality drop. Supports dimension reduction to cut vector-database cost. Same no-upgrade-path caveat: moving off it requires re-embedding everything.

AZURE FLAG CORRECTED this pass: voyage-4 IS in Azure AI Foundry. The catalog lists voyage-4-embedding-model, voyage-4-lite-embedding-model, voyage-4-large-embedding-model and voyage-context-4, with an Azure Marketplace SaaS offer from MongoDB (Voyage AI's owner). January 2026 generation. ~$0.06 per 1M for voyage-4, $0.12 for voyage-4-large, $0.02 for voyage-4-lite; 200M free tokens per account on the v4 generation and 33% off via Batch API. Supports 2048/1024/512/256 dimensions with a shared embedding space across the v4 series.

From Cohere. Model id 'embed-v-4-0' confirmed in the Microsoft Learn 'Foundry Models sold by Azure' list and in the ai.azure.com catalog; available both as serverless pay-as-you-go and on Managed Compute. Turns text, images and interleaved text-and-image into a single vector space. I confirmed availability and capability but not current per-token Azure pricing. Migration note from Microsoft Q&A: the Azure AI Inference beta SDK retired 2026-08-26 — use the OpenAI /v1-compatible API or Cohere's own SDK pointed at your Azure endpoint.

Verified on ai.google.dev. $0.15 per 1M tokens, text only. The generally-available fallback if Gemini Embedding 2's preview status is a blocker. NOT in Azure Foundry.