Anthropic calls for slowdown, ships Opus 5.5

Ten days after Anthropic CEO Dario Amodei published "We Must Pace the Frontier," calling on the industry to slow down AI development, the company released Claude Opus 5.5 on September 22, 2026 (UTC-4). It's the first model in a new Claude 5.5 family, the first release since the pacing call, and — awkwardly for the slowdown narrative — it beats GPT-6 Astra on most benchmarks at a fraction of the cost.

What shipped

Opus 5.5 is built for long-running agentic coding and knowledge work. It keeps the 1M token context window and 128K max output from Opus 5, but the pricing drops to $4/$20 per million input/output tokens — 20% cheaper than Opus 5's $5/$25. Cache reads fall to $0.20 per million, a 60% cut. Anthropic says typical workloads cost 40% less to run, and output generation is over 30% faster.

Four breaking changes affect code migrating from Opus 5: thinking can't be disabled, forced tool use returns an error, and two other API behavior shifts are detailed in the platform docs. Fast mode remains available at $8/$40 for up to 2.5x speed.

The model was tested before release by external evaluators including Frontier Design and METR. On Anthropic's automated behavioral audit — its most comprehensive alignment test — Opus 5.5 is the strongest-performing model the company has ever tested. It's also more resistant to prompt injection than Opus 5, tying Fable 5.1 for the lowest success rate on a Gray Swan benchmark.

The benchmark picture

Anthropic's own comparison table shows Opus 5.5 leading six of eight published benchmarks, including a clean sweep of agentic coding.

Source: Anthropic official announcement, September 22, 2026. GPT-6 Astra and GPT-5.6 Sol figures as reported by OpenAI. "—" indicates the model was not included in that benchmark by either publisher.

The two benchmarks where Astra wins — AutomationBench (41.4% vs 40.0%) and Terminal-Bench-Science (64.6% vs 58.7%) — are narrow gaps. On coding, the story is more lopsided: Opus 5.5 beats Astra's top Terminal-Bench score by 8.5 points and matches it for roughly 40% of the cost per task. On FrontierCode, Opus 5.5 at default effort beats Astra at max effort for about a fifth of the cost.

Artificial Analysis separately placed Opus 5.5 first overall, with an 81/100 composite score ranking #4 of 234 models and the #1 slot in agentic work.

Why it matters

The efficiency story here is bigger than the benchmark story. Opus 5.5 costs 40% less to run than Opus 5 while outperforming it on every measured dimension. That's not a minor tweak — it's a structural shift in what frontier intelligence costs. If Anthropic can sustain this trajectory, the "frontier model = $50+/M output tokens" pricing assumption that anchored the industry breaks. GPT-6 Astra at $10/$50 suddenly looks expensive against a model that beats it on coding for a fifth the cost.

The timing is the headline. Amodei's September 12 essay argued that frontier labs should coordinate to slow development, warning that uncontrolled scaling poses catastrophic risks. Ten days later, Anthropic shipped a model that beats its own flagship on most benchmarks and undercuts it on price. You can argue this is exactly what "pacing" looks like — a measured, safety-audited release rather than a reckless sprint — but the optics are terrible. Competitors and critics will ask: if pacing means releasing a better, cheaper model every ten days, what does speeding up look like?

Early tester quotes tell the real-world story. GitHub reported Opus 5.5 solved more terminal tasks than Opus 5 in less than half the steps. Clio ran it unattended for 18 hours across six repositories. Quantium cut a 38-prompt, four-day task down to 11 prompts in three hours. Optiver matched Opus 5 quality in half the turns and tokens, cutting workload cost by 40-50%. These aren't benchmark games — they're production workloads getting dramatically cheaper.

The critical lens

Let's not ignore the contradiction. Amodei's pacing essay was explicit: "We believe the frontier should be paced." The antitrust lawsuit filed six days later (Buist v. Anthropic PBC, September 18) alleges that Amodei, Altman, Musk, and Hassabis conspired to slow AI development in violation of the Sherman Act. Now Anthropic releases Opus 5.5 — a model that, by its own benchmarks, outperforms GPT-6 Astra on six of eight tests. If the industry is being "paced," it's pacing upward at a remarkable clip.

The benchmark methodology deserves scrutiny too. Anthropic runs its own comparison table, and the GDPval-AA score of 1846 for Opus 5.5 vs. 1542 for Astra is a 304-point gap — larger than the typical gap between model generations. Artificial Analysis's independent ranking is more reassuring, but users should wait for third-party replications before treating these numbers as settled. The AutomationBench result (40.0% vs Astra's 41.4%) was run by Zapier without fallback models, meaning safeguard interventions counted as failures — a methodology that disadvantages models with stricter safety guardrails.

And the "40% cheaper" claim needs unpacking. It's 40% cheaper on typical workloads at default settings, driven partly by lower per-token prices and partly by Opus 5.5 using fewer tokens per task. The per-token price drop is only 20%. The efficiency gains are real but uneven — workloads with heavy cache usage benefit most (60% cheaper cache reads), while short, simple tasks see less advantage.

What to watch

Sonnet 5.5 and Haiku 5.5 are confirmed for the coming weeks. If they carry similar efficiency gains, Anthropic will have a full product line that undercuts competitors at every tier — not just the high end. Watch for OpenAI's response: GPT-6 Sol, rumored for this week, now has a harder bar to clear. If Sol ships at $10/$50 with benchmark parity to Opus 5.5, the pricing pressure becomes acute.

Watch the antitrust case. Opus 5.5's release undercuts the "conspiracy to slow down" narrative — if Anthropic is releasing frontier-beating models ten days after calling for a slowdown, plaintiffs will struggle to show coordinated suppression. But the pacing essay itself remains evidence in the case, and the timing won't help Anthropic's optics.

Watch enterprise migration. At $4/$20 with Terminal-Bench leadership, Opus 5.5 is the first model in months that gives CTOs a clear reason to switch from GPT-6 Astra. If Cursor and GitHub Copilot usage shifts toward Claude in the next 4-6 weeks, the price war has begun in earnest.