AI Model Release Tracker 2026
Last updated: August 5, 2026, 18:00 UTC · Updated within 72 hours of any qualifying release · Next scheduled review: August 12, 2026 | Maintained by: Axis Intelligence Research & Sarah Mitchell
Eleven frontier model launch events between April 24 and August 3, 2026 are logged below, each traced to the developer’s own announcement, release note or API changelog. Seven of the eleven — 63.6% — did not reach unrestricted general availability on their announcement date. Axis Intelligence Research measures that delay as the Announcement-to-Availability Lag.
Quick Answer: What Is the AI Model Release Tracker?
The Axis Intelligence Research AI Model Release Tracker is a permanently maintained log of frontier AI model launches, with one entry per launch event, each carrying a primary-source link fetched at entry time. As of August 5, 2026 it holds 11 launch events plus 1 availability event. Across the 7 launches whose availability gate has closed, the Axis Announcement-to-Availability Lag (AAL™) has a baseline median of 0 days and a mean of 7.1 days. Four launches remain gated.
Key Findings
According to Axis Intelligence Research, 7 of 11 frontier model launch events tracked between April 24 and August 3, 2026 shipped behind a preview, partner or export restriction rather than into unrestricted general availability.
Axis Intelligence Research finds the baseline Announcement-to-Availability Lag across the 7 closed launches in the tracker is a median of 0 days and a mean of 7.1 days, as of August 5, 2026.
According to Anthropic’s June 9, 2026 announcement, Claude Fable 5 launched with classifiers that route cybersecurity, biology, chemistry and distillation queries to Claude Opus 4.8, and Anthropic states these trigger in under 5% of sessions.
According to OpenAI’s July 9, 2026 release post, GPT-5.6 is priced per million tokens at $5 input and $30 output for the Sol tier, following a limited preview that began June 26, 2026.
According to DeepSeek’s April 24, 2026 release note, DeepSeek-V4-Pro carries 1.6T total parameters with 49B active and DeepSeek-V4-Flash 284B total with 13B active, both open-weight with a 1M-token default context.
How We Curate This List
The inclusion rule is narrow on purpose. A row exists only if a developer’s own announcement, release note, model card or API changelog was fetched and read at entry time, and the URL and retrieval date were logged. Aggregator timelines, leaderboard sites and launch-tracker databases are used to find candidate events. They are never the source of a dated row.
What counts as an entry:
- A new named model or model family made available to external users by the organization that trained it — general availability, public preview, limited partner release, or open-weight publication.
- A material change to an existing model’s availability: suspension, restoration, tier migration, or a shift from preview to production under a new model ID.
- Open-weight publication of a model previously available only through an API.
What does not count:
- Roadmap statements, keynote mentions, and executive comments about models with no shipped artifact. The one exception is a model announced and then measurably not shipped, which is logged with
release_type = announced_not_shippedprecisely because the gap is the finding. - Fine-tunes, distills and community re-uploads of another lab’s weights.
- Version aliases and endpoint redirects that expose no new capability.
- Pricing changes, deprecations and shutdowns. These are tracked in the maintenance protocol as review triggers, not as entries.
Where dates come from. The announcement_date is the date on the developer’s own post or changelog entry. The availability_date is the date the model became callable by any paying customer without a partner agreement, waitlist, verified-identity gate or geographic restriction. Where those are the same date, the lag is zero. Where availability has not been reached, the row stays open and the day count runs to the snapshot date.
What we leave out, and why we say so. Three fields in the current dataset are deliberately empty. DeepSeek publishes its V4 pricing as an image inside the release note, so the price columns for that row carry no values pending a fetch of the pricing page. OpenAI’s GPT-5.6 page notes a July 30 price reduction for the Terra and Luna tiers without stating whether the prices listed further down predate or follow it, so only the Sol tier is recorded. Alibaba’s Qwen blog renders client-side and returned no server-side content, so the Qwen3.8-Max preview date is flagged as secondary-sourced and is not used as a verified figure in this article. An empty cell with a stated reason is more useful to a researcher than a filled cell with an unstated guess.
Axis Announcement-to-Availability Lag (AAL™): Methodology and Baseline Reading
AAL™ v1.0 — baseline reading as of August 5, 2026.
AAL measures, in days, the interval between a model’s first public announcement by its developer and the date it becomes generally available to any paying customer without a gate. A gate is a partner program, a waitlist, an identity-verification requirement, a geographic restriction, or an export control.
Formula. For each launch event i:
AAL_i = availability_date_i − announcement_date_i
Where availability_date is empty, the row is open and the reported figure is snapshot_date − announcement_date, which is a floor, not a final value. Open rows are excluded from the median and mean. Availability events — suspensions and restorations that are not themselves launches — carry no AAL.
Inputs. All 12 rows of the tracker CSV. Every announcement_date and availability_date is drawn from a developer-published document listed in the row’s source_url.
Baseline reading. Seven closed launches: 0, 0, 0, 0, 13, 15, 22 days. Median 0 days. Mean 7.1 days. Four open launches at 6, 36, 57 and 78 days as of the snapshot date. This is the inaugural reading and describes no history, because none has been computed.
What the median hides, which is the point. A median of zero would suggest models ship the day they are announced. Four of them did — three Gemini Flash-tier releases and DeepSeek V4, all of which appeared in a changelog on the day they became callable. The distribution is bimodal, not centred: either a model ships immediately or it sits behind a gate for weeks. The mean of 7.1 days is the more honest summary of the second group’s drag on the whole.
Derived reading — gated-launch share. Of 11 launch events, 7 carried a non-zero lag: 63.6% as of August 5, 2026.
Limitations. AAL treats a same-day changelog entry as an announcement, which flatters labs that announce and ship in one motion and penalizes labs that preview publicly. It cannot see private previews. It does not weight by model capability, so a Flash-tier point release and a frontier flagship count once each. And an announcement date recovered from a keynote rather than a document is weaker evidence than a changelog line — which is why the one row in that position, Gemini 3.5 Pro, carries a verification flag.
AI Model Releases 2026: The Tracker
Reverse chronological. Every entry links to the developer’s own document.
August 3, 2026 — Alibaba releases Qwen3.8-Max
Alibaba’s Qwen team moved its flagship from a preview endpoint to a production model ID, with API access priced at $2 per million input tokens and $6 per million output. Open weights were promised rather than published; at the time of entry no Alibaba model card for Qwen3.8-Max had appeared on Hugging Face. That promise is the row’s review trigger.
Access: general API · AAL: 15 days (preview date flagged) · Source: Alibaba (Qwen), launch announcement, official Qwen account Verification flag: the qwen.ai blog page renders client-side and returned no fetchable text. The July 19 preview date and the parameter and context figures circulating in coverage are secondary-sourced and are not stated here as verified.
July 30, 2026 — Google puts Gemini Robotics ER 2 into public preview
Two embodied-reasoning endpoints, one of them optimized for real-time streaming through the Live API. Both accept text, image, video and audio, and support function calling with blocking behaviour for physical robot actions — the detail that separates a robotics endpoint from a multimodal chat model.
Access: public preview · AAL: 6 days, open · Source: Google, Gemini API release notes
July 21, 2026 — Gemini 3.6 Flash and Gemini 3.5 Flash-Lite reach general availability
Google shipped both stable. It describes 3.6 Flash as more token-efficient than 3.5 Flash at a lower price, framing the change as a response to developer complaints about output verbosity — a rare case of a release note naming the feedback it answers. Flash-Lite is positioned as a subagent tier for high-volume automation. The same entry deprecates the temperature, top_p and top_k sampling parameters.
Access: general availability, both · AAL: 0 days, both · Source: Google, Gemini API release notes
July 9, 2026 — OpenAI ships the GPT-5.6 family
Sol, Terra and Luna reached general availability across ChatGPT, Codex and the API, thirteen days after a limited preview restricted to a small group of partners. OpenAI reports Sol at 80 on the Artificial Analysis Coding Agent Index and 92.2% on BrowseComp under its ultra multi-agent setting, and describes its cyber safeguards as blocking roughly ten times more potentially harmful activity than previous models. Sol is priced at $5 input and $30 output per million tokens.
Access: general availability · AAL: 13 days · Source: OpenAI, GPT-5.6 Verification flag: a July 30 update note on the same page records price reductions for Terra and Luna. The pricing paragraph does not state whether its figures precede or follow that change, so only the Sol prices are recorded in the dataset.
June 30, 2026 — Gemini Omni Flash enters public preview
A multimodal model for high-speed video generation and conversational video editing, generating short clips at 720p from text or animating still images, then editing them in dialogue. Still preview-gated at the snapshot date.
Access: public preview · AAL: 36 days, open · Source: Google, Gemini API release notes
June 12, 2026 — Anthropic suspends access to Claude Fable 5 and Mythos 5
Three days after launch, both Mythos-class models were pulled to comply with a US export-control directive. Access was restored on July 1. The entry is logged as an availability event rather than a launch: nothing shipped, but the availability of a shipped model changed, and the tracker exists to record exactly that.
Access: suspended, subsequently restored · Source: Anthropic, availability statement
June 9, 2026 — Anthropic launches Claude Fable 5 and Claude Mythos 5
The first Mythos-class model released for general use, alongside a partner-restricted sibling. Anthropic’s own framing is unusually specific about the trade-off it made: Fable 5 ships with classifiers that hand cybersecurity, biology, chemistry and distillation requests to Claude Opus 4.8 instead of answering, tuned conservatively enough to catch some benign requests, triggering in under 5% of sessions on Anthropic’s early data. Both models are priced at $10 input and $50 output per million tokens. Mythos 5 is the same underlying model with cyber safeguards lifted, available only through Project Glasswing.
Access: Fable 5 general, restored July 1 · Mythos 5 partner-only · AAL: Fable 5, 22 days · Mythos 5, 57 days and open · Source: Anthropic, Claude Fable 5 and Claude Mythos 5
May 19, 2026 — Gemini 3.5 Flash reaches general availability
Released stable and made the model behind the gemini-flash-latest alias, which is the operationally significant half of the announcement: alias reassignment moves traffic without any customer action.
Access: general availability · AAL: 0 days · Source: Google, Gemini API release notes
May 19, 2026 — Gemini 3.5 Pro announced, not shipped
Logged as announced_not_shipped. The Gemini API release notes run through July 30, 2026 without a gemini-3.5-pro entry, while sibling models in the same family — 3.5 Flash, 3.5 Flash-Lite, 3.6 Flash — all have dated GA lines. The absence in the primary changelog is the evidence, and it is more reliable than the presence of a launch date in coverage.
Access: not available · AAL: 78 days, open · Source: Google, Gemini API release notes Verification flag: the May 19 announcement date is attributed to Google I/O 2026 by secondary coverage. Google’s own I/O post names Gemini 3.5 Flash. Primary confirmation of the 3.5 Pro announcement date is pending; the day count above should be read as provisional.
April 24, 2026 — DeepSeek ships V4-Pro and V4-Flash with open weights
Two models on one day, with API access and open weights simultaneously: V4-Pro at 1.6T total parameters with 49B active, V4-Flash at 284B total with 13B active. DeepSeek made a 1M-token context the default across its official services and routed the legacy deepseek-chat and deepseek-reasoner aliases to V4-Flash ahead of their retirement on July 24, 2026. The release is still labelled a preview.
Access: general API and open weights · AAL: 0 days · Source: DeepSeek, V4 Preview Release Verification flag: V4 pricing appears on the release note as an image. Price fields are empty pending a fetch of the pricing page. <details> <summary><strong>Archive — entries before April 24, 2026</strong></summary>
The archive opens with the first monthly rollover. Entries are moved here, never deleted, and retain their original entry IDs and source URLs. Archived rows remain in the downloadable dataset. </details>
What the 2026 Release Pattern Actually Shows
Analysis by Sarah Mitchell, AI/ML Editor.
The number that matters in this dataset is not a benchmark. It is 63.6% — the share of tracked launches that did not reach unrestricted availability on announcement day.
Read the four zero-lag rows and a pattern appears: three of them are Google Flash-tier models that surfaced as a changelog line on the day they became callable, and the fourth is DeepSeek publishing weights and an API endpoint together. None of the four is a frontier flagship. Every flagship in the tracker — Fable 5, Mythos 5, GPT-5.6, Qwen3.8-Max — arrived through a gate of some kind. Two of those gates were governmental.
That is a change in how the top of the market ships, and it is worth naming precisely rather than dramatically. A preview period is not new; labs have staged rollouts for years. What is new is the reason stated in the primary documents. OpenAI describes previewing GPT-5.6 to the US government ahead of launch. Anthropic launched Fable 5 with a classifier that hands a defined set of topics to a weaker model and published the fallback rate. Then an export-control directive took both Mythos-class models offline for eighteen days. When two of the three largest labs publish safety and regulatory reasoning as part of the availability announcement itself, the gate has stopped being a capacity-management detail.
The counter-reading deserves space, because it is plausible. Four zero-lag launches in fourteen weeks says the routine cadence is intact and fast, and a tracker that counts flagships and Flash-tier point releases as one event each will over-weight the drama. Qwen3.8-Max’s fifteen days was a product preview, not a regulator. And 11 events is a small sample — small enough that one quarter of open-weight releases could move the gated share ten points in either direction.
The honest position at baseline: the lag exists, it clusters on the frontier tier, and it is measurable. Whether it is widening is a question this tracker will be able to answer in about six months, and cannot answer today. The one thing we can already say is that the Gemini 3.5 Pro row — 78 days from announcement to nothing, evidenced by silence in Google’s own changelog — is the kind of gap that a release-date roundup, assembled from coverage rather than documents, will report as a shipped model.
Methodology
Every row originates from a document fetched during a research session, with the URL and retrieval date recorded in the dataset. Searches locate candidate URLs; they never supply figures. Where a developer publishes a figure only as an image, or where a page renders client-side and returns no fetchable text, the field is left empty and the reason is recorded in the row’s verification_flag.
The Announcement-to-Availability Lag is computed as described above from the announcement_date and availability_date columns. The median, mean and gated-launch share published in this article recompute exactly from the CSV; no figure in the prose exists outside it.
Known limitations. The tracker’s coverage begins April 24, 2026 and is being backfilled. It over-represents organizations that publish structured changelogs — Google’s API release notes make Gemini events trivially loggable, while labs that announce through social posts and press briefings are systematically harder to enter, which biases the dataset toward the better-documented rather than the more significant. Regional models with no English-language primary source are under-represented and are being added as sources are verified.
About This Dataset
ai-model-release-tracker.csv contains every tracked entry with full provenance: source_org, source_document, source_url, retrieved_date, is_primary, axis_calculated, method_note and verification_flag on each row. Temporal coverage begins April 24, 2026 and extends to the snapshot date. Licensed CC BY 4.0.
Citation: Axis Intelligence Research, AI Model Release Tracker, 2026.
Frequently Asked Questions
How many AI models were released in 2026?
The Axis Intelligence Research tracker logs 11 frontier model launch events between April 24 and August 3, 2026, each traced to a developer-published document. That count is deliberately narrower than aggregator totals running into the hundreds, because it excludes fine-tunes, community re-uploads, endpoint aliases and any release whose primary source we could not fetch.
What was the most recent frontier AI model release?
As of August 5, 2026, the most recent entry is Qwen3.8-Max, released by Alibaba’s Qwen team on August 3, 2026 to general API access at $2 input and $6 output per million tokens, with open weights promised but not yet published.
What is the Announcement-to-Availability Lag?
It is an Axis Intelligence Research metric measuring the days between a developer’s first public announcement of a model and the date that model becomes callable by any paying customer without a preview, partner, identity or export gate. Its baseline reading across 7 closed launches is a median of 0 days and a mean of 7.1 days.
Which 2026 model launches were restricted at release?
Seven of eleven tracked launches carried a restriction: Claude Fable 5 and Claude Mythos 5, GPT-5.6, Qwen3.8-Max, Gemini Omni Flash, Gemini Robotics ER 2, and Gemini 3.5 Pro, which was announced and has not shipped.
How often is this tracker updated?
Within 72 hours of any qualifying release, with a scheduled review every Wednesday and a monthly summary drawn computationally from this tracker’s own dataset. Every update bumps the page’s modification date and adds a dated line to the changelog in the maintenance protocol.
Why do some entries have empty fields?
Because the figure could not be verified from a document we fetched. Each empty field carries a stated reason in the row’s verification_flag. Three currently apply: DeepSeek’s V4 pricing is published as an image, OpenAI’s GPT-5.6 page does not disambiguate pre- and post-reduction prices for two tiers, and Alibaba’s Qwen blog returns no server-side content.
Review trigger: a new frontier model launch, an availability change to any tracked model, or the closing of any open AAL row. Whichever comes first.
Go Deeper
- EU AI Act Compliance Tracker — primary-source enforcement log with the Axis Intelligence EU AI Act Enforcement Gap Ratio™
- AI Copyright Lawsuits Tracker — 24+ cases across the US, UK, and Germany with the Bartz Settlement Efficiency Ratio™
- EU AI Act Compliance Checker — interactive tool for assessing your organization’s regulatory exposure