Data · Models

AI model release timeline, 2022 to 2026

Every dated launch in one table, with context window, list price, licence and API availability — and an explicit note on what is verified and what is still missing.

P
Plugsky Data desk
Baseline 20 November 2026 · reviewed 20 November 2026

Short answer. The timeline starts with text-davinci-003 and ChatGPT in late 2022, moves through the GPT-4 and Claude 2 launches in 2023, the multimodal and open-weight wave of 2024, the reasoning and million-token-context releases of 2025, and the 2026 generation that is still being priced. This edition carries 41 rows, each with a primary source.

Baseline: 20 November 2026. We publish the rows we can verify, not a target count. The schema supports 400+ models and the file is updated weekly; a model appears only when its release date, context window or licence can be read from the provider or the model card.

What a row contains. Each row answers five questions a developer actually asks when choosing a model: when did it ship, how much context does it take, what does it cost, can I download the weights, and is there an API. The filters in the JSON file mirror those columns, so you can select “open weights, at least 128K context, under $1 per million input tokens” without parsing prose.

FieldTypeNotes
release_dateISO datePublic availability, not the paper date
modelstringFamily and size, e.g. Llama 3.1 405B
providerstringDeveloper, not the host
contextintegerMaximum input tokens at launch
input_usd_mtoknumber or nullNull for open weights with no list price
output_usd_mtoknumber or nullStandard tier, launch price
licencestringProprietary, Apache 2.0, MIT, Llama community
apibooleanWhether a hosted API existed at launch

2022–2023: the API era

DateModelProviderContextIn / outLicenceAPI
2022-11-28text-davinci-003OpenAI4,097$20 / $20ProprietaryYes
2022-11-30ChatGPT (research preview)OpenAI4,096n/aProprietaryNo
2023-03-14GPT-4 (8K)OpenAI8,192$30 / $60ProprietaryYes
2023-06-13GPT-3.5 Turbo 16KOpenAI16,385$3 / $4ProprietaryYes
2023-07-11Claude 2Anthropic100,000$8 / $24ProprietaryYes
2023-07-18Llama 2Meta4,096Self-hostLlama 2No
2023-09-27Mistral 7BMistral8,192Self-hostApache 2.0Yes
2023-11-06GPT-4 TurboOpenAI128,000$10 / $30ProprietaryYes
2023-11-21Claude 2.1Anthropic200,000$8 / $24ProprietaryYes
2023-12-06Gemini 1.0 ProGoogle32,768$0.50 / $1.50ProprietaryYes
2023-12-11Mixtral 8x7BMistral32,768Self-hostApache 2.0Yes

2024: multimodal and open weights

DateModelProviderContextIn / outLicenceAPI
2024-02-15Gemini 1.5 ProGoogle1,000,000$3.50 / $10.50ProprietaryYes
2024-03-04Claude 3 HaikuAnthropic200,000$0.25 / $1.25ProprietaryYes
2024-03-04Claude 3 OpusAnthropic200,000$15 / $75ProprietaryYes
2024-04-18Llama 3 70BMeta8,192Self-hostLlama 3No
2024-05-13GPT-4oOpenAI128,000$5 / $15ProprietaryYes
2024-06-20Claude 3.5 SonnetAnthropic200,000$3 / $15ProprietaryYes
2024-07-18GPT-4o miniOpenAI128,000$0.15 / $0.60ProprietaryYes
2024-07-23Llama 3.1 405BMeta128,000Self-hostLlama 3.1No
2024-09-12o1-previewOpenAI128,000$15 / $60ProprietaryYes
2024-12-11Gemini 2.0 FlashGoogle1,000,000$0.10 / $0.40ProprietaryYes
2024-12-26DeepSeek V3DeepSeek128,000$0.27 / $1.10Open weightsYes

2025: reasoning and long context

DateModelProviderContextIn / outLicenceAPI
2025-01-20DeepSeek R1DeepSeek128,000$0.55 / $2.19MITYes
2025-01-31o3-miniOpenAI200,000$1.10 / $4.40ProprietaryYes
2025-02-24Claude 3.7 SonnetAnthropic200,000$3 / $15ProprietaryYes
2025-03-25Gemini 2.5 ProGoogle1,000,000$1.25 / $10ProprietaryYes
2025-04-05Llama 4 ScoutMeta10,000,000Self-hostLlama 4No
2025-04-14GPT-4.1OpenAI1,000,000$2 / $8ProprietaryYes
2025-05-22Claude Opus 4Anthropic200,000$15 / $75ProprietaryYes
2025-08-07GPT-5OpenAI400,000$1.25 / $10ProprietaryYes
2025-09-29Claude Sonnet 4.5Anthropic200,000$3 / $15ProprietaryYes

2026: the current generation

Only rows with a verifiable date or a current listing appear here. Models listed without a launch date use the first day we observed them and are flagged date_confidence: observed, not presented as launch facts.

DateModelProviderContextIn / outLicenceAPI
2026-03-17GPT-5.4OpenAI272K short$2.50 / $15ProprietaryYes
2026-03-17GPT-5.4 nanoOpenAI272K short$0.20 / $1.25ProprietaryYes
2026-04-23GPT-5.5OpenAI272K short$5 / $30ProprietaryYes
2026-09-03GPT-6 AstraOpenAI272K short$10 / $50ProprietaryYes
2026-09-29GPT-6.1 SolOpenAI272K short$2 / $10ProprietaryYes
2026-11-20GPT-6.2 AstraOpenAI272K short$8 / $40ProprietaryYes
2026-11-20Claude Fable 5.2AnthropicSee model page$8 / $40ProprietaryYes
2026-11-20Claude Opus 5.5AnthropicSee model page$4 / $20ProprietaryYes
2026-11-20Claude Sonnet 5.5AnthropicSee model page$2 / $10ProprietaryYes
2026-11-20Claude Haiku 5.5AnthropicSee model page$0.10 / $0.50ProprietaryYes

Two patterns: the top slot costs more in absolute terms than it did in 2023, and the cheap slot is roughly 1/100th of it. “Which model?” is now a routing question.

Filters, JSON and ICS

The JSON file is an array of objects with the fields above. The ICS calendar carries the same dates as all-day events, so a release calendar can be imported into a team calendar without a script. Both are regenerated weekly, on Monday, and the page’s dateModified is the generation date.

{
  "release_date": "2024-05-13",
  "model": "GPT-4o",
  "provider": "OpenAI",
  "context": 128000,
  "input_usd_mtok": 5.0,
  "output_usd_mtok": 15.0,
  "licence": "proprietary",
  "api": true,
  "source": "https://openai.com/index/hello-gpt-4o/",
  "date_confidence": "primary"
}

Filter examples: open weights with 100K+ context, API models under $1 per million input tokens, or one quarter’s releases. The file is deliberately flat so a spreadsheet, a jq one-liner and a build script read it the same way.

Methodology and gaps

  • Release date. The date the model became generally available through an API or a public weights release. Announcement dates and paper dates are not used.
  • Context. Maximum input at launch. A later extension is a new row only if the provider treats it as a new model.
  • Price. Standard tier at launch, USD per 1M tokens. Open-weight models show “self-host” because there is no list price.
  • Licence. The weights licence, not the API terms. “Open weights” without an OSI licence is labelled as such.

The gap that matters: 41 rows is not 400. Coverage is strongest for OpenAI, Anthropic and Google and weakest for Chinese providers and small open-weight labs. A missing row means unverified, not unreleased.

Sources

Links checked 20 November 2026. We cite the primary page for each figure. Where a provider publishes a price behind a login, we say so and do not guess.

FAQ

How many models are in the timeline?

This edition contains 41 verified rows from November 2022 to November 2026. The schema supports more than 400 and the file is regenerated weekly, but a model is only added when its release date, context window or licence can be read from a primary source.

Does the timeline include open-weight models?

Yes. Llama 2, Llama 3, Llama 3.1, Llama 4, Mistral 7B, Mixtral 8x7B and DeepSeek R1 are included with their weights licence, and the price column reads “self-host” because there is no list price.

What counts as a release date?

General availability through an API or a public weights release. Announcement posts, paper submissions and waitlist openings are not used, because they do not tell you when you could call the model.

Can I import the dates into a calendar?

Yes. The ICS export carries every dated row as an all-day event, and the JSON export carries the same rows with context, price, licence and API availability so you can filter before importing.

Why are some 2026 rows dated 20 November 2026?

Because the provider lists the model and its price but not a launch date. Those rows use the first date we observed them and are flagged as observed rather than primary, so a reader can tell a launch fact from a snapshot.

Use the data, then test it on your own workload

Plugsky publishes its plan caps and model limits. Start free with no card, or compare the flat plans against per-token billing.

Start Free → See plans