Anthropic launches Claude Opus 4.5, claims state of the art for coding and agents
Anthropic on November 24 released Claude Opus 4.5, its new flagship model, available the same day in the Claude consumer apps, on the Claude Developer Platform under the model identifier claude-opus-4-5-20251101, and through Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Azure AI Foundry. Reuters, CNBC and Bloomberg covered the launch as a competitive move against OpenAI's GPT-5.1 family and Google's Gemini 3, with Anthropic positioning Opus 4.5 specifically against coding, agentic, and computer-use workloads.
The release sits inside a compressed 2025 cadence. Anthropic shipped Claude Sonnet 4.5 on September 29 with chief product officer Mike Krieger framing it as the new default for most coding work, while Opus 4.1 had remained reserved for the highest-stakes tasks at $15/$75 per million tokens. The new $5/$25 price effectively brings Opus-class capability into Sonnet-tier economics — a roughly two-thirds reduction. Context window holds at 200,000 tokens, with new beta features for memory and context editing aimed at long-running agent sessions.
In its launch blog post, Anthropic wrote: "Our newest model, Claude Opus 4.5, is available today. It's intelligent, efficient, and the best model in the world for coding, agents, and computer use." The company added: "Opus 4.5 is a step forward in what AI systems can do, and a preview of larger changes to how work gets done." No individual executive was quoted in the post itself; Reuters and CNBC published the announcement without a fresh on-record quote from CEO Dario Amodei or Mike Krieger, attributing positioning statements to Anthropic as a corporate source.
On benchmarks, Anthropic reports 80.9 percent on SWE-bench Verified — the first publicly disclosed score above 80 percent on that test — alongside 59.3 percent on Terminal-Bench 2.0 and a range of 88.9 to 98.2 percent on tau2-bench. The company also published efficiency comparisons, claiming that at medium effort the new model matches Claude Sonnet 4.5 on SWE-bench Verified while using roughly 76 percent fewer output tokens, and that at high effort it exceeds Sonnet 4.5 by 4.3 percentage points using about 48 percent fewer tokens. Reuters, citing Anthropic's recent reporting cycle, noted Anthropic's annualised revenue run-rate had climbed past 7 billion US dollars by late October.
Industry reaction across customers cited by Anthropic — including GitHub, Cursor, Notion, Lovable, and Warp — emphasised the price-to-capability shift more than raw benchmark numbers. More cautious analysts flagged two issues: that the 80.9 percent SWE-bench result still depends on harness configuration and parallel test-time compute methods footnoted in Anthropic's announcement, and that effort-control parameters complicate apples-to-apples comparison with OpenAI and Google models. Concerns over agentic safety remained: Anthropic's own system card describes Opus 4.5 as the most aligned frontier model the company has shipped, while The Guardian and other outlets continued to flag broader agentic-AI deployment risks.
For us at Enpo Sekai, the relevant signal is not the SWE-bench number — it is the price. At 5 dollars in and 25 dollars out per million tokens, Opus-class quality moves into the cost band where it can plausibly run sustained character and persona pipelines, not just one-off enterprise coding sessions. That changes our latitude on long-running NPC dialogue, multi-turn voice agents, and on-device-plus-cloud hybrid characters, where token budgets used to push us toward Sonnet-tier models. The architectural moat we care about — character voice, persona consistency, local-first runtime — does not move; the unit economics underneath it do.
