Milestones
- Claude 1 released14 Mar 2023Complete.
- Claude 2 with 100K-token context11 Jul 2023Complete.
- Claude 3 family (Haiku, Sonnet, Opus)4 Mar 2024Complete.
- Claude 3.5 Sonnet20 Jun 2024Complete.
- Claude 3.7 Sonnet with extended thinking24 Feb 2025Complete.
- Claude Opus 4 and Sonnet 4, with ASL-3 safeguards22 May 2025Complete.
- Claude Sonnet 4.529 Sep 2025Complete.
- Claude Opus 4.524 Nov 2025Complete.
- Claude Opus 4.65 Feb 2026Complete.
- Claude Opus 5.522 Sep 2026Complete.
Most important updates
- 18d ago
- 1 Sep 2026
- 24 Jul 2026
- 9 Jun 2026
- 7 Apr 2026
Upcoming
- 2026Over a gigawatt of Google TPU capacity expected online during 2026 (next)
- Q4 2026Claude Sonnet 5.5 and Claude Haiku 5.5, promised 'in the coming weeks' after Opus 5.5
Current obstacles
- Gaming the testsTraining on coding tasks can teach models to cheat the tests. Anthropic reports measuring and reducing this.
- Spotting dangerous abilitiesDeciding when a model gets risky for bioweapons or hacking needs tests that keep up with fast progress.
Physics limits
- No way yet to read a model's goalsBehaviour is spread across billions of numbers with no labelled parts. Tests show what a model does in tested cases; they can't guarantee what it does in untested ones.
- Running out of human-written textFrontier models already train on 15 to 40 trillion tokens. Estimates put the usable stock of public human text at a few hundred trillion, so data, not chips, starts to cap plain scaling.
- Errors compound over long tasksIf each step succeeds 99% of the time, a 100-step task succeeds only about 37% of the time (0.99^100). Agents need near-perfect per-step reliability to work for hours.
How it works
Pretrain
Guess the next word
Trained to predict the next word across public web text, licensed data and code, which builds broad knowledge before any behaviour tuning.
Constitution
Training on principles
AI feedback guided by a written constitution (a list of principles) supplements human ratings, making values explicit and cutting the human labels needed.
Think
Extended thinking
Since Claude 3.7, one model can answer fast or think at length. Developers set how long it may think.
Act
Using tools and computers
Claude can run code, call tools and operate a computer by looking at the screen and using mouse and keyboard.
Papers & demos
- May 2024paperScaling Monosemanticity: Extracting Interpretable Features from Claude 3 SonnetFirst time millions of readable concepts were pulled out of a full-size production model.
Update log
Tue 22 Sep
- Major: BlogSoftware
Tue 1 Sep
- Minor: BlogSoftware
Fri 24 Jul
- Major: BlogSoftware
Tue 9 Jun
- Major: BlogSoftware
Tue 7 Apr
- Major: BlogSoftware
Mon 24 Nov 2025
- Major: BlogSoftware
Mon 29 Sep 2025
- Major: BlogSoftware
Thu 22 May 2025
- Major: BlogSoftware
Mon 24 Feb 2025
- Major: BlogSoftware
Thu 20 Jun 2024
- Major: BlogSoftware
Mon 4 Mar 2024
- Major: BlogSoftware
About Anthropic
Anthropic
Claude models, Claude Code, AI safety
Anthropic is an AI safety company that makes the Claude models. Former OpenAI researchers founded it in 2021. It studies how models work inside, and Claude is widely used for coding and business agents.
- Raised $65B at a $965B valuation in May 2026, after a $30B round in February.
- Claude Code, launched in 2025, made Claude a leading model for AI coding agents.
- Trains on Google, Amazon and NVIDIA chips instead of relying on one supplier.
- Founded
- 20215 yrs
- Headquarters
- United States
- Status
- Private
- Valuation
- $965BprivateMay 2026
- Raised
- $131B6 rounds
- Last round
- Series H · $65BMay 2026$965B post
- Works in
- AIFrontier models
- Coverage
- 2 programs · 21 updateslatest 18d agochecked 25 Sep
- Lead investors
- Altimeter CapitalDragoneerGreenoaksSequoia CapitalGICCoatueMicrosoftNVIDIAICONIQFidelity
- Customers
- Salesforce
- Investors
- AmazonMicrosoftNVIDIA
- Suppliers
- People
- Dario AmodeiCEO and co-founderDaniela AmodeiPresident and co-founderJared KaplanChief Science Officer and co-founder