Research scope: This study counts releases listed in the official API changelogs of OpenAI, Anthropic, Google, DeepSeek and Mistral, read on November 10, 2026. Each date is the first day a new text model name appeared in the API (experimental, preview or generally available), January 2024 to October 2026. It measures how often models ship, not their quality or price, and counts depend on how each developer names and bundles models.
Short answer: Across 11 full quarters from January 2024 to September 2026, five developers released 135 new text models in their public APIs. Counting January to September only, releases rose from 23 models in 2024 to 44 in 2025, then held at 40 in 2026. We call the main pattern the Decimal Clock: the median gap between new version numbers fell from 112 to 63 days, while whole-number generations still arrived 308 to 496 days apart.
What did we count?
We audited text model releases across five provider changelogs on November 10, 2026. The primary sources include the OpenAI API changelog, the Anthropic Claude Platform release notes, the Google Gemini API release notes, the DeepSeek API news page, and the Mistral changelog. For the Claude 3 family, we included Anthropic's launch announcements from March 4 and March 13, 2024, which predate its formal platform release notes. Each row in our dataset represents a distinct text model name, defined as a new family, a new version number, or a new tier, on the exact date it first surfaced in the public API.
We excluded several categories to keep the count focused strictly on distinct text model releases. Our count omits dated snapshots of an existing name, such as gpt-4o-2024-08-06 or DeepSeek-V3-0324, along with generic latest aliases. It leaves out models restricted to approved organizations, tool-specific releases for computer use, deep research, or search, and third-party models hosted by a provider. We also excluded audio, image, video, embedding, moderation, and OCR models.
Across the 11 quarters from Q1 2024 to Q3 2026, the five developers released 135 text models across 101 combined launch days. Mistral released 42 models across 28 launch days, while OpenAI published 40 models across 26 launch days. Anthropic introduced 22 models across 20 launch days, Google added 19 models across 17 launch days, and DeepSeek delivered 12 models across 10 launch days.
Developers differ in how many model sizes they drop at a single time. OpenAI averaged 1.54 models per launch day, closely matched by Mistral at 1.50 models per launch day. DeepSeek averaged 1.20 models per launch event, Google posted 1.12, and Anthropic averaged 1.10.
Finding 1: Releases doubled in 2025 and moved to the US big three in 2026
Total model releases across the five developers rose through 2024 and 2025 and stayed high in 2026. The count ran at 5 models in Q1 2024, 7 in Q2 2024, 11 in Q3 2024, and 9 in Q4 2024. Volume accelerated to 12 models in Q1 2025, 16 in Q2 2025, and 16 in Q3 2025, reaching a peak of 19 models in Q4 2025, the busiest quarter in the dataset. Activity settled into 2026 with 12 models in Q1, 11 in Q2, and 17 in Q3.
Comparing the identical nine-month stretch from January to September across calendar years shows this expansion clearly. The five providers published 23 models during the first nine months of 2024, 44 models in the same period of 2025, and 40 models in the first nine months of 2026. While the overall volume in 2026 matched the pace set in 2025, the source of those releases shifted heavily toward OpenAI, Anthropic, and Google.
OpenAI, Anthropic and Google increased their output in each of these nine-month periods. Together, OpenAI, Anthropic, and Google shipped 11 models between January and September 2024, 24 models in that span in 2025, and 32 models in 2026. OpenAI went from 4 models to 14 and then 15; Anthropic scaled from 4 to 5 and then 10; Google increased from 3 to 5 and then 7.
DeepSeek and Mistral moved the other way or stayed low. Mistral went from 10 models in the first nine months of 2024 to 17 in 2025, but dropped to 4 in 2026. DeepSeek held steady at lower volumes, moving from 2 models in 2024 to 3 in 2025 and 4 in 2026. The shift was clearest in Q3 2026, when OpenAI, Anthropic, and Google accounted for 15 of the 17 total releases, compared to just 6 of 16 models in Q3 2025. Mistral listed zero new text models in Q3 2026 before introducing Mistral Large 4 in preview on October 6, 2026.
New text model names by the quarter they first appeared in each developer's official API changelog. Snapshots, aliases, restricted-access and non-text models are not counted.
Finding 2: The Decimal Clock
Our primary finding is the Decimal Clock: on the main model line of each developer, the median time between new version numbers (for example GPT-5.4 to GPT-5.5) fell from 112 days for the 19 gaps that ended in 2024 to 2025 to 63 days for the 21 gaps that ended from January to October 2026, while whole-number generations (for example GPT-5 to GPT-6) still arrived 308 to 496 days apart. We track this metric along each lab's primary text sequence: OpenAI GPT, Anthropic Claude, Google Gemini, DeepSeek V, and Mistral Large. Shorter gaps mean more point releases between whole-number generations.
From January to October 2026, OpenAI, Anthropic, and Google each shipped six new version numbers along their primary lines. OpenAI shipped GPT-5.3, 5.4, 5.5, 5.6, 6, and 6.1. Anthropic published Claude 4.6, 4.7, 4.8, 5, 5.1, and 5.5. Google deployed Gemini 3.1, 3.5, 3.6, 3.7, 3.8, and 4.
This tempo represents a sharp break from 2025. Across all twelve months of 2025, OpenAI shipped five main-line versions, Anthropic shipped four, and Google released two. Those 2025 releases included OpenAI GPT-4.5, 4.1, 5, 5.1, and 5.2, Anthropic Claude 3.7, 4, 4.1, and 4.5, and Google Gemini 2.5 and 3.
Comparing gaps that closed in 2024 to 2025 against those closing between January and October 2026 shows how the cadence contracted. OpenAI saw its median interval between version updates fall from 98 days to 53 days. Anthropic shortened its median gap from 87 days to 56 days. Google registered the most dramatic acceleration, dropping its median gap from 238 days to 52 days. Across all five providers, the median time separating version numbers dropped from 112 days to 63 days.
The fastest updates in 2026 arrived in a matter of days. OpenAI moved from GPT-5.3 on February 24 to GPT-5.4 on March 5, a gap of just 9 days. Anthropic moved from Claude Opus 4.8 on May 28 to Claude Fable 5 on June 9, an interval of 12 days. Google advanced from Gemini 3.7 Flash on August 13 to Gemini 3.8 Flash on September 2, taking 20 days.
DeepSeek and Mistral operated on a more extended schedule. DeepSeek recorded a median gap of 112 days across 3 version intervals in 2024 to 2025, followed by 173 days across 2 intervals in 2026. Mistral Large posted a median gap of 149 days across 3 intervals in 2024 to 2025, followed by 308 days for its single gap in 2026. These samples are small.
Main lines only: OpenAI GPT, Anthropic Claude, Google Gemini, DeepSeek V and Mistral Large. Each dot is the first API date of a new version number.
Median days between consecutive new version numbers on each main line, grouped by the year the gap ended. DeepSeek and Mistral rest on 1 to 3 gaps.
Finding 3: Whole-number generations still take about a year
While point releases sped up in 2026, whole-number generations stayed far apart. The median interval across all eight whole-number model transitions in the dataset was 387.5 days. Developers iterated on decimal increments at high frequency, but whole-number versions still arrived roughly ten to sixteen months apart.
The three longest transitions took well over a year. Mistral Large 2 to Mistral Large 3 took 496 days, running from July 24, 2024 to December 2, 2025. DeepSeek V3 to DeepSeek V4 took 484 days, spanning December 26, 2024 to April 24, 2026. Claude 3 to Claude 4 required 444 days, stretching from March 4, 2024 to May 22, 2025.
Generational transitions inside OpenAI and Anthropic clustered close to the one-year mark. GPT-5 to GPT-6 took 392 days, running from August 7, 2025 to September 3, 2026. Claude 4 to Claude 5 arrived after 383 days, dating from May 22, 2025 to June 9, 2026.
The three shortest transitions belong to Google and Mistral. Gemini 2.0 to Gemini 3 took 342 days between December 11, 2024 and November 18, 2025, while Gemini 3 to Gemini 4 arrived 329 days later on October 13, 2026. Mistral Large 3 to Mistral Large 4 recorded 308 days between December 2, 2025 and October 6, 2026. In 2026, the decimals sped up while the whole numbers stayed near one year apart.
Days between the first API release of consecutive whole-number versions on each main line.
Finding 4: Google moved to generally available Flash releases
Google changed how new models first appear in its API. All 12 Google models first listed in 2024 and 2025 appeared as experimental or preview releases. In 2026, 6 of Google's 8 new models launched with general availability on the day they first appeared, starting with Gemini 3.5 Flash on May 19, 2026.
Most of Google's 2026 releases were Flash models: 6 of its 8 new models were Flash or Flash-Lite. The company released Gemini 3.5 Flash on May 19, Gemini 3.6 Flash on July 21, Gemini 3.7 Flash on August 13, and Gemini 3.8 Flash on September 2. That cadence delivered four successive Flash versions within 106 days.
After Google listed Gemini 3.1 Pro Preview on February 19, 2026, it did not introduce another non-Flash Gemini model until Gemini 4 Argon arrived on October 13, 2026.
Across all five developers, the proportion of models launching without general availability held relatively stable. Non-GA models represented 21.9% of total releases in 2024, 19.0% in 2025, and 18.2% in 2026. These figures include Mistral Labs models and DeepSeek's temporary V3.2-Speciale, which both counted as not generally available at launch.
Finding 5: New tier names replace size words
In 2026, OpenAI, Anthropic, Google and DeepSeek each introduced tier names that were not in their 2024 and 2025 lineups. OpenAI released its final mini and nano models within our observation window on March 17, 2026, with GPT-5.4 mini and GPT-5.4 nano. When OpenAI launched GPT-5.6 on July 9, 2026, it introduced Sol, Terra, and Luna, and followed with Astra alongside GPT-6 on September 3, 2026. Exactly 8 of the 16 models OpenAI published between January and October 2026 used these new naming conventions.
Anthropic also broadened its tiered naming structure beyond its established sizes. On June 9, 2026, the developer introduced Claude Fable 5, creating a fourth tier alongside Opus, Sonnet, and Haiku. It expanded that tier on September 1, 2026, with the introduction of Claude Fable 5.1.
Other providers also drifted away from conventional naming schemes. Google dropped the Pro and Flash designations for its primary generation launch, introducing Gemini 4 as Gemini 4 Argon on October 13, 2026. DeepSeek V4, launched on April 24, 2026, came as two models: V4 Pro and V4 Flash.
Mistral preserved its established tier naming of Large, Medium, Small, and Ministral while releasing specialized models across specific disciplines. Across Mistral's 43 total rows, 9 models targeted software development under the Codestral and Devstral lines. Another 6 rows represented reasoning models under the Magistral label, and 3 rows covered mathematics through the Mathstral and Leanstral series.
What this means for teams that build on these APIs
With version gaps at the big three sitting at a median of 53 to 56 days, an engineering team pinned to specific model identifiers encounters a new version roughly every two months from each provider. Managing updates can no longer be handled as an annual project. Engineering groups must treat prompt evaluation and model migration as continuous routines rather than rare maintenance events.
Teams should keep model identifiers stored strictly in configuration files and environment variables rather than hardcoded inside application logic. Decoupling the codebase from explicit model tags reduces migration friction when newer decimals land. A gateway such as Tokely lets you switch the model name without changing the integration.
New tier names like Sol, Terra, Luna, Fable, Argon, and Flash require independent testing against existing production workloads. Teams cannot rely on legacy size words or familiar brand tiers to deduce performance characteristics, latency profiles, or pricing structures. Direct evaluation against your own task distributions is necessary before routing production traffic to newly released model tiers.
Methodology and limitations
We compiled this dataset by reviewing official changelogs on November 10, 2026, verifying each entry against its public source URL on that date. Each release date reflects the timestamp assigned by the provider in its official changelog. These entry dates can differ by a day or so from blog announcements.
Release counts depend heavily on how individual developers package, brand, and bundle their models. A lab that introduces multiple sizes on a single day logs multiple models in this dataset, which is why launch days should be considered alongside raw model counts. In addition, the cadence medians calculated for DeepSeek and Mistral rest on small samples of 1 to 3 gaps, meaning individual release decisions sway their metrics significantly.
Our analysis deliberately omits providers and models that do not meet strict documentation criteria. We excluded xAI because its release notes list updates by month rather than specific calendar days and do not list all earlier Grok models. OpenAI's gpt-oss open-weight models do not appear in the OpenAI API changelog and are not tracked here.
Our version gap calculations track main lines only, covering GPT, Claude, Gemini, DeepSeek V, and Mistral Large. While OpenAI's o-series and Mistral's specialized coding, math, and reasoning lines appear in the raw dataset, we excluded them from the median gap figures. This study measures API release velocity over time and does not measure output quality, inference latency, or pricing changes.
Download the data
One CSV with 139 rows contains the raw data across the observation window, covering developer, date, quarter, model ID, line, version, tier, stage, source URL, and verification date. Cite as: Tokely, "How Often Do AI Labs Ship New Models? Release Cadence of OpenAI, Anthropic, Google, DeepSeek and Mistral (2024-2026)", November 2026, https://tokely.me/blog/ai-model-release-cadence-2026.
Originally published at tokely.me.
Top comments (0)