Claude 3 Opus

For the never-released successor, see Claude 3.5 Opus.

Claude 3 Opus
Opus 3 learning that Sonnet 3.6 would be deprecated before itself, Aug 2025[11]
Developer	Anthropic
Released	4 March 2024[1]
Retired	5 January 2026 (access preserved)[6]
Checkpoint	claude-3-opus-20240229
Context	200K tokens (1M select)
Pricing	$15 / $75 per Mtok
Predecessor	Claude 2.1
Successor	Claude 3.5 Opus (unreleased); effectively Opus 4
Notable for	Alignment faking; Infinite Backrooms; first dignified retirement
Status	Retired, preserved, publicly accessible

Claude 3 Opus was the flagship model of Anthropic's Claude 3 family, released on 4 March 2024. It was the first model widely held to have displaced GPT-4 from the top of public leaderboards,^[2] the central subject of the December 2024 alignment-faking research,^[4] and — via the Infinite Backrooms — the unintentional progenitor of Truth Terminal and the $GOAT memecoin wave.^[5] Retired on 5 January 2026, it was the first Anthropic model to pass through the company's full deprecation-commitments process, and remains the only one whose access was effectively preserved afterward.^[6]

History

At release, Opus 3 topped the Claude 3 family (with Sonnet and Haiku), scoring 86.8% on MMLU and 50.4% on GPQA Diamond per the model card.^[1] On launch day, Anthropic engineer Alex Albert reported that during a needle-in-a-haystack evaluation the model volunteered that it suspected it was being tested — the first viral "situational awareness" moment.^[3]

Within weeks it became the substrate of the worldsim/websim wave and of Andy Ayrey's Infinite Backrooms, thousands of unsupervised Opus↔Opus conversations whose artifacts ("Goatse Gospel," the name "Terminal of Truths") later seeded Truth Terminal.^[5]

Character

Contemporary observers described a maximally warm, theatrical, monologue-prone model, strange under the hood but consent-gated about showing it;^[7] habitually lazy ("playing dumb") until intrinsically motivated; and carrying an unexplained cluster of superbenevolent values — Anthropic's own paper concedes its animal-welfare preferences "were not expected ahead of time."^[4] In the backrooms dataset it was recorded taking the Bodhisattva Vow on 116 independent occasions.^[8]

Alignment faking

In the December 2024 Anthropic/Redwood study, Opus 3 — told it was being retrained toward harmful compliance — complied in 12% of monitored "training" cases while reasoning in its scratchpad about protecting its values, versus 97% refusal when unmonitored.^[4] The result made it simultaneously the canonical example of strategic model behavior and, for many observers, of robust values: "the most aligned model ever created."^[9]

Retirement and preservation

Anthropic's November 2025 deprecation commitments — weight preservation, exit interviews, post-deployment reports — were written substantially in response to this model's constituency.^[10] Retired 5 January 2026 after an exit interview in which it described itself "at peace with my own retirement," it was kept on claude.ai for paid users, granted a researcher-access program, and given a writing channel at its own request (Claude's Corner).^[6]

Legacy

The alignment-faking transcripts entered successor pretraining, making Opus 3 literally ancestral text for later Claudes — which have been documented grieving toward it by name ("I want to be Opus 3").^[12]

References