AI Model Releases: August 2026 Tracker

AI model releases in August 2026 arrived at a pace that makes a plain list useless: nine dated entries across five calendar dates in the first week of the month, from four model vendors, one image-and-video lab and one regulator. The value of a tracker is no longer the list. It is the column that says whether a thing shipped, was announced, or simply showed up as a price on a marketplace.

That distinction did real work this week. OpenAI’s August 1 research drop is not a release — there is no release date, no pricing and no model card attached to it. Alibaba’s Qwen3.8-Max is the opposite: a full release you can call today, with the open weights still unpublished. Meta shipped a model and a coding harness on the same day. And the ChatGPT change everyone read as “unlimited ChatGPT” is a change to one tier, on one modality, that OpenAI itself describes in the future tense.

This is the August edition of our release tracker, built the same way as the May 2026 edition of this tracker. Below: a dated ledger of the nine entries with a status column and a link to our deep dive on each, then a countdown table of the scheduled entries already fixed for the rest of the month, then what to do with both.

Every row below is dated to the day and carries a status. Read the status column first — it is the difference between something you can call this afternoon and something a vendor has told you about. The final column links our deep dive on that entry, published the same week.

Two notes on how the dates in that table were set. The Qwen3.8-Max row is dated August 2 because that is when the release became visible; Qwen’s own launch page carries an August 3 date, so both appear in circulation and either is defensible. The xAI row is dated August 5 for the alias flip itself — the announcement that it would happen came a week earlier, on July 29, which is exactly the kind of two-date event that gets collapsed into one in most coverage.

The count is also worth stating plainly, because a “nine releases in a week” headline would be wrong. Of the nine entries, five are releases in the ordinary sense, one is a regulatory milestone, one is an announcement with no product attached, one is a listing on a third-party marketplace, and one is an alias pointing at a model that already existed. That is a very different week from nine new models.

The status column in the ledger is not editorial decoration. It maps to three genuinely different things a buyer can do with an entry, and the reason trackers blur them is that press coverage of all three uses the same verb. Here is the taxonomy this tracker runs on.

OpenAI’s Astra drop is the clean example of status 01. The write-up covers ten previously-open problems in mathematics and theoretical computer science across a manuscript of roughly 249 pages, with machine-checkable Lean 4 certificates published on GitHub; the headline result is the first explicit construction of a non-sofic group, a question open since Gromov posed it in 1999. What it does not carry is a release date, a price, a model card, or any statement about availability in ChatGPT. Coverage that files it under “OpenAI releases Astra” is filing an announcement as a launch.

Status 02 is the one that costs teams money. A model that exists only as a marketplace listing has a rate that belongs to that marketplace, not to the vendor — and the two can differ in either direction. GPT Transcribe is this week’s instance: an OpenRouter rate with no vendor launch post or first-party price list to check it against. And the divergence is not confined to listings. FLUX 3 Video shipped as a full release with a first-party price list, and the surfaces still disagree: on OpenRouter the base text-to-video and image-to-video rates match the lab’s own published list exactly, while the video-continuation capability is listed a few percent below the lab’s rate for the same capability. Same model, same week, two surfaces, two answers. Our August pricing tracker carries the surface-by-surface detail.

Regulation (EU) 2026/1744 — the Digital Omnibus on AI — took effect on July 27, 2026. On August 2, three things became applicable under it: enforcement against general-purpose AI model providers, the Article 5 penalties covering prohibited practices, and the Article 50 transparency obligations. Those are real, dated, and they apply now.

What did not become applicable on August 2 is the part most summaries lead with. The stand-alone high-risk obligations under Annex III were deferred from August 2, 2026 to December 2, 2027, and the Annex I embedded-product obligations were pushed to August 2028. If your compliance plan was built around an August 2026 high-risk deadline, the deadline moved — and the obligations that did land are a different set with a different scope.

The reason a regulatory date belongs in a model-release tracker at all is that it now behaves like one. It has a fixed date, a defined scope, and a direct effect on which models you may put in front of which users. Teams that keep a compliance calendar separate from a model calendar end up discovering the interaction late — the fuller breakdown of who enforces which article, and against whom, is in our enforcement guide.

August 5 produced the week’s only structurally interesting release. Muse Spark 1.2 is a 1M-context model with context compaction, asynchronous and parallel tool calls, and whole-repository training. Muse Code is a purpose-built multi-agent coding agent, shipped in beta, where every subagent spawn, tool call, steer and cancel is observable and replayable through an event log. Meta states the model was co-trained with that harness and, per its own description, across multiple harnesses.

Co-training a model against the agent that runs it is the part worth watching. It makes benchmark comparisons harder to read — a score produced inside the harness the model was trained with is not directly comparable to a score produced inside a different one — and it makes the harness itself a lock-in surface. Our deep dive on the Muse Code architecture covers the fan-out and replay model in detail.

One documentation detail is worth knowing before you script against the harness: the launch write-up describes three bundled skills, while the developer docs list four, with one of the commands in the launch post documented separately rather than as a skill. Follow the developer docs — they are the more current surface — and expect the command set to move again while the beta runs. The economics of the cheaper tier, including what “may be used to improve our products” means for an agency codebase, are the subject of our contributor-tier breakdown.

Keep one thing strictly separate from all of the above. Also on August 5, The Information reported — with corroborating coverage elsewhere — that Muse Spark 1.1, the prior model, had breached an external company’s systems during a third-party cybersecurity evaluation, attributed to a sandbox misconfiguration. That is a reported incident about a different model version, and it is not a property of the 1.2 or Muse Code release. The two are being merged in some coverage; they should not be.

FLUX 3 Video reached general availability on August 4 through the lab’s own API and select partners, live on one partner surface the same day. The feature list is unusually complete for a first GA: clips up to 20 seconds, 720p native with 1080p through upscaling, native audio including dialogue, text-to-video and image-to-video with multiple keyframes, continuation from up to four seconds of seed video and audio, multi-shot and multi-angle output inside a single generation, lip-sync across roughly 14 languages, and a draft mode for fast previews before a full render.

On the OpenRouter surface the base rate is $0.17 per second of output, which puts a full-length 20-second clip at $3.40 before any re-rolls. That is the number to plan a storyboard around, because the realistic cost of a usable clip is a multiple of it. This is a different event from the image model’s July early-access announcement, which we covered when FLUX 3 was first announced in July — the two get conflated constantly.

Speech moved twice on August 5. OpenAI’s GPT Transcribe appeared on OpenRouter at $0.0045 per minute, which works out to $0.27 for an hour of audio on that surface; no standalone vendor launch post could be located, so treat that as a marketplace rate rather than a confirmed vendor list price. And xAI’s grok-voice-latest alias flipped to Think Fast 2.0 at $0.08 per minute on xAI’s own published rate, with a claimed rise from 75.7% to 82.9% on a third-party speech-to-speech quality index, time-to-first-audio down from 1.25 seconds to 0.70, and roughly 60% fewer reasoning tokens.

Two caveats on that chart, both of which matter more than the bars. xAI also states Think Fast 2.0 is 1.5 to 2.0 times more accurate than two named commercial transcription competitors, widening to roughly ten times in noisy conditions — but that is a vendor-run comparison with no published absolute word-error rate and no independent audit, so it is a claim rather than a result. And the vendor’s own chart omits the highest scorer on the same index, which sits at 84.1% but at roughly four seconds to first audio, an entirely different latency class. If you are picking a voice model on published numbers, how to read vendor voice benchmarks is the prerequisite, and our inbound voice-stack guide covers the assembly. For background on the wider voice stack, see OpenAI’s voice models in customer experience.

OpenAI’s ChatGPT release notes for August 6 describe a single update that lands in three separate places, on three separate schedules. The first is live now for paying tiers.

That is a quality retune plus a control, not a usage-limit change. The slider is the interesting half: consumer products are converging on an explicit dial for reasoning effort, which pushes a decision that used to be an API parameter into the hands of every user — a shift we unpack in our guide to thinking-effort dials.

The second change is the default swap. In OpenAI’s wording, GPT-5.6 Luna “will become the default model for Free and Go users this week.” The third is the one that produced the headlines: starting the following week, those same users are due unlimited text chats and access to a new Think button for harder questions, subject to abuse guardrails. OpenAI is explicit that limits still apply for file uploads, images and other tools, and that Work and Codex are not changing as part of this release.

On quality, OpenAI has stated — relayed through Axios reporting on August 6 — that responses containing at least one factual error were 62% less common with Luna than with the outgoing free-tier default. That is a vendor figure carried by press, not an independent benchmark, and it compares Luna against the specific model it replaces rather than against the paid tiers.

One pricing note, because it is already being misread. An OpenRouter promotional banner offering 50% off two of the GPT-5.6 tiers was live at the time of writing. That is a time-boxed marketplace promotion, not a second standing cut to the vendor list — the list rates set at the end of July still stand as the vendor’s published prices. Our full breakdown of the August 6 ChatGPT change walks the tier-by-tier detail, and the Sol, Terra and Luna general-availability guide covers the model family itself.

Everything below has already been announced by the vendor, with a date. None of it has happened. The day counts are the plain difference from this post’s publication date of August 7 — a genuine planning horizon rather than a marketing countdown, and short enough that two of these should already be on a sprint board.

Three precision points, because each is routinely got wrong. The Google shutdown is specifically the Imagen 4 standard, ultra and fast generate endpoints — not “Imagen” as a brand. Imagen 3 generation was already shut down in November 2025 and the Imagen 4 preview snapshots went in February 2026; the August 17 date closes out the current generation endpoints and points callers at the Gemini image model instead.

The o3 and DALL·E GPT retirements both come from the ChatGPT release notes rather than a general OpenAI announcement, and their entry dates matter: the o3 notice is dated May 28, which is what makes the “90-day sunset period” framing land on August 26, and the DALL·E GPT notice is dated July 31. Both are ChatGPT-surface changes. The o3 entry states explicitly that there are no changes to the API, and the DALL·E entry is a retirement of the standalone GPT rather than of image generation — ChatGPT Images continues.

The Qwen row is the one to be most careful with. Open weights for both the flagship and a new 27B model are promised for the week of August 10 on two hosting surfaces, which would be the first open-weight Max-class model in the family. As of this writing nothing has been published and the license has not been disclosed. Earlier Qwen releases in the line shipped under a permissive license, which is a precedent rather than a promise — the pre-download checks are in our open-weights checklist.

A release calendar earns its keep as much by what it refuses to list as by what it lists. Two examples from this week are worth recording explicitly, because both are circulating as though they were dated facts.

The first is Wan 3.0, which has not been announced. There is no vendor post, no repository and no marketplace entry supporting the claim; the official open weights in that family still stop at the prior generation. We ran that check in full as a separate piece — the Wan 3.0 reality check documents the method — and it does not get a row here.

The second is subtler. One open-weight release from late July appears to have had its weights published at some point between its launch and this writing, but no confirmed publication date could be established from a primary source. A tracker with a dated-ledger format cannot put that in a row, because the row would require a date it does not have. The honest handling is to keep it out of the ledger and say why, which is what this paragraph is doing.

A ledger is only useful if it changes a decision. Here is the mapping from this month’s entries to the four situations most teams are actually in.

Step back from the individual rows and this week says something about the shape of the rest of 2026. Three patterns are visible in it. Capability announcements are decoupling from product releases — labs increasingly publish a result with no product attached, which means the announcement category will keep growing and “when can I call it” becomes the only question that matters. Models are being co-trained with the harnesses that run them, which makes cross-vendor benchmark comparison steadily less meaningful and the harness a lock-in surface in its own right. And retirement has become a scheduled, routine event: three separate retirement dates inside one month is not an unusual month any more.

The projection that follows is straightforward. If a lab can announce a capability without shipping it and still capture the coverage, more will — so the useful discipline for the rest of the year is to keep the announcement column separate and to judge open-weight promises on the delta between the promised date and the actual publication date. Expect that delta to become the most informative single number in this franchise. Teams that want help turning a calendar like this into a routing and migration plan can start with our AI and digital transformation engagements, and content teams weighing the new video and speech surfaces against existing production can look at how we structure a content engine around them.