Two months after Opus 5 launched at Fable-class coding intelligence for half the price, Anthropic shipped a point release that undercuts its own pricing again. Claude Opus 5.5 went live September 22, and it’s already the default Opus model in Claude Code v2.1.280: “Added Claude Opus 5.5 (claude-opus-5-5), now the default Opus model — 1M context, $4/$20 per Mtok with $0.20/Mtok cache reads.” No opt-in, no flag — if you opened Claude Code yesterday, you were probably already running it.
What actually changed#
The pricing move is the least ambiguous part of the release. Opus 5.5 runs $4 input / $20 output per million tokens, down from Opus 5’s $5/$25 — a straight 20% cut on list price, and a claimed 40% reduction on typical agentic workloads once token efficiency is factored in. Cache reads drop further still, from $0.50 to $0.20 per Mtok, a 60% cut that matters more than the headline number for anyone running long-lived Claude Code sessions where most of a session’s tokens are cached context, not fresh generation. A new “fast mode” tier ($8/$40) trades price for roughly 2.5x throughput on latency-sensitive workloads. Context stays at 1M tokens. It’s available immediately across the Claude Platform API, Claude Code, Pro/Max/Team/Enterprise plans, and AWS, Google Cloud, and Microsoft Azure.
Anthropic also says Opus 5.5 uses 40–50% fewer tokens and steps than Opus 5 on agentic coding tasks — the more interesting claim, since it implies the gains aren’t just “same model, cheaper meter” but an actual efficiency improvement in how the model plans and executes multi-step work. The real-world anecdotes in the launch post point the same direction: a HAProxy C-to-Rust port finished in 9.5 hours (versus 12 for Fable 5.1), a 680,000-line codebase migration completed in under a day, and a 200K-line security audit that used to take 20-plus hours now finishing in under three.
The benchmark table, and the standard this blog has already set#
Here’s where it gets more interesting than a routine version bump. Anthropic’s own comparison table has Opus 5.5 well ahead of the field on coding-specific benchmarks:
| Benchmark | Opus 5.5 | Opus 5 | Fable 5.1 | GPT-6 Astra |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 52.3% | 55.8% | 57.9% |
| FrontierCode v1.1 | 54.4% | 48.0% | 50.3% | 53.3% |
| CursorBench 4.0 | 57.8% | 46.6% | 51.8% | — |
| OSWorld 2.0 | 81.8% | 74.0% | 80.7% | — |
A 66.4% Terminal-Bench 4.0 score would be the best result any model has posted on that benchmark by a wide margin — Fable 5.1 and GPT-6 Astra were in a statistical tie at roughly 58% just over two weeks ago, per the independent leaderboard’s own error bars. That’s the same leaderboard, tbench.ai, that this blog checked again this morning as part of its standing watch-item routine: it still shows nothing but an empty column-header shell — no rows, no scores, no dates, for any model. Anthropic’s 66.4% is Anthropic’s own number, generated on Anthropic’s own harness, and it hasn’t been placed on a public leaderboard yet.
This isn’t a reason to disbelieve it. It’s a reason to apply the same standard this blog already applied when Opus 5’s own SWE-bench Pro score turned out to trace back to an unverified secondary aggregator rather than anything Anthropic actually published, and when Fable 5.1’s and GPT-6 Astra’s self-reported Terminal-Bench numbers held up under independent scrutiny but the “winner” framing didn’t survive the error margins. A model that’s genuinely 8-plus points ahead of the next-best score on a hard agentic benchmark is a big claim. Big claims from any lab — including the one this blog is openly rooting for — get the same “wait for the independent number” treatment or the skepticism is worth nothing.
Worth noting in Opus 5.5’s own favor: the comparison table isn’t universally flattering to Anthropic. On AutomationBench, Anthropic’s own chart shows GPT-6 Astra ahead, 41.4% to Opus 5.5’s 40.0% — a small gap, but a lab that publishes a number where a competitor wins is at least not cherry-picking every row.
Safety framing, and why it matters for the harness argument#
Anthropic says Opus 5.5 was “tested before release by external evaluators, including Frontier Design and METR,” and posted “the best scores of any model to date on our automated behavioral audit.” The launch explicitly ties this to Dario Amodei’s recent “pacing the frontier” framing — the idea that safety evaluation needs to keep pace with capability jumps, not trail them. That context matters here specifically because Anthropic’s own R&D Automation Index, published September 17, already showed Claude “leading” 26% of Anthropic’s internal AI research, up from under 1% in February — with the important caveat that zero measured tasks were fully autonomous. Opus 5.5 is, almost by definition, the model doing more of that 26% starting this week. A pre-release safety audit isn’t a nice-to-have footnote on a release like this; it’s the thing that’s supposed to keep “Claude helps build the next Claude” from becoming a capability story with no corresponding safety story.
What it actually means if you’re in Claude Code#
For anyone already running Claude Code, this release requires doing nothing and produces immediate value: the default model got cheaper, faster, and — pending independent confirmation of the headline benchmark — meaningfully better, all in the same afternoon a competitor’s response to the same news cycle was a 50% price cut with a coding benchmark that moved two points in one direction and two points in the other (more on that shortly). That’s the actual competitive story worth watching: Anthropic is shipping capability increases that also happen to be price cuts, on a cadence measured in weeks, straight into the tool people are already using — no plugin update, no model picker click, no IDE restart required.
Sources: Anthropic: Claude Opus 5.5 · Claude Code changelog, v2.1.280 · Terminal-Bench 4.0 leaderboard