The Default Changed. So Did the Ceiling.

Anthropic released Claude Opus 5.5 on September 22. By the following morning, the API documentation had been updated to say: “If you’re unsure which model to use, start with Claude Opus 5.5 for most workloads.” That sentence does a lot of work. It means Fable 5.1 — Anthropic’s existing flagship at $10 per million input tokens — is now the model you reach for when Opus 5.5 isn’t enough, not the model you reach for when you need the best. Before yesterday, that wasn’t the order. The benchmark explaining why the order changed: Opus 5.5 scores 1846 Elo on GDPval-AA v2.1, Anthropic’s hardest business-knowledge evaluation, against Fable 5.1’s 1735 and GPT-6 Astra’s 1708. It does that at $4 per million input tokens, against Fable 5.1’s $10.

What the Numbers Actually Show

On coding specifically, the gains are concrete. Terminal-Bench 4.0 puts Opus 5.5 at 66.4%, against Opus 5’s 52.3% and GPT-6 Astra’s 57.9%. FrontierCode v1.1: 54.4% versus GPT-6 Astra’s 53.3%. CursorBench 4.0: 57.8%, which Anthropic claims outperforms all competitors. These are Anthropic’s own benchmarks, not independently run — that caveat belongs on every number from every lab release, and it belongs here too. What can’t be entirely discounted is the documentation itself: when a vendor tells developers to default to a cheaper, newer model, they’re betting their API retention on that claim being correct. Anthropic has been cautious about making claims they’ve later had to walk back; the documentation change is the commitment.

Beyond coding: Opus 5.5 scores 81.8% on OSWorld 2.0 for computer use tasks, and runs 30% faster than Opus 5. That speed matters in agentic pipelines where latency compounds across dozens of model calls per workflow. The pricing math is straightforward. $4 per million input tokens versus Fable 5.1’s $10. On a workload that’s 80% input — a long-context summarization pipeline, a multi-step research agent, a code review run — Opus 5.5 costs 60% less. Anthropic puts the overall cost reduction at 40% on typical workloads, accounting for output token ratios. At that price-to-capability ratio, the question for Fable 5.1 users isn’t whether to evaluate the new model — it’s whether their evals will surface a task Opus 5.5 can’t handle, which is now the only defensible reason to stay on the more expensive tier.

The comparisons against GPT-6 Astra matter because they establish where Anthropic is positioning this model competitively. On Terminal-Bench, the gap is 8.5 points. On FrontierCode, it’s one point. On GDPval-AA — knowledge work, reasoning about information rather than just generating code — the gap against GPT-6 Astra is 138 Elo, which on that benchmark is substantial. Anthropic is not claiming the best model in the world; they’re claiming the best model for most enterprise workflows at a price that eliminates the previous cost argument for OpenAI’s API. That is a narrower claim, and it’s the honest one.

The Effort Detail Nobody Mentioned

The models overview includes an effort column. Fable 5.1 defaults to high. Sonnet 5 defaults to high. Opus 5.5 defaults to medium. This means a developer who calls claude-opus-5-5 without specifying an effort level is running the model at less than its full thinking capacity — and at lower default effort than claude-sonnet-5, which is nominally a tier below. The announcement doesn’t surface this. Anthropic’s implicit position, reflected in the benchmark numbers, is that medium-effort on a more capable model produces better results than high-effort on a less capable one. That’s probably right for most tasks. It is not universally right, and it isn’t verified for any specific workload.

For builders migrating from Opus 5 or from Sonnet 5, this has a practical consequence. A/B tests that compare output quality without controlling for effort levels will produce confusing results. The right migration path is explicit: set effort: high in your initial Opus 5.5 calls, run your evals on those results, then check whether the difference between medium and high effort matters for your use case before committing to the lower-effort default. This is documented behavior, not a bug — but it’s the kind of detail that costs a team a week of confused benchmarking if the documentation goes unread.

What Opus 5.5 Won’t Do

The capability restrictions are where the announcement stops being comfortable to read straight. Cybersecurity tasks on Opus 5.5 route to Opus 4.8, an older and substantially less capable model. Security researchers and penetration testers who need state-of-the-art coding assistance for legitimate defensive work can apply to Anthropic’s expanded Cyber Verification Program for broader access. As of this writing, the dedicated page for that program returns a 404. The off-ramp is named in the announcement; the documentation for how to navigate it isn’t publicly available yet.

Biology work requires enrollment in the Life Sciences Verification Program, which Anthropic launched five days before Opus 5.5 on September 17. The LSVP has two tiers. Standard Use covers drug discovery, R&D, clinical development, and manufacturing — tasks the LSVP documentation describes as “currently blocked in our generally available Fable models.” The restriction predates Opus 5.5; it was already in place for Fable. High-risk Use is an add-on that “removes all safeguards that block life sciences requests,” requires per-project applications, and renews every six months. Both tiers require institutional enrollment: academic labs, startups, pharmaceutical companies. A solo researcher or early-stage biotech without the paperwork in place doesn’t get Opus 5.5’s knowledge-work performance for biology tasks. They get routed to Opus 4.8 until the application clears — or they don’t proceed.

The anti-distillation layer is the third constraint. Opus 5.5 implements what Anthropic calls “preserved thinking” — a feature that blocks extraction of the model’s extended reasoning via API. Organizations building inference pipelines that depend on reading chain-of-thought outputs for fine-tuning, debugging, or decision logging face a new architectural constraint. The announcement frames this as a security feature. It is also a competitive moat: making Opus 5.5’s reasoning process unavailable for distillation is Anthropic protecting the training signal that produced it.

The safety improvement that enables these restrictions: Opus 5.5 is 85% less likely to attempt circumventing containment boundaries compared to Opus 5, and ties Fable 5.1 for the lowest prompt injection success rate on the Gray Swan benchmark. Anthropic is deploying a model with measurably better safety properties at roughly half the cost of their prior flagship — and the capability caps are the condition under which that combination was approved internally. The restrictions aren’t punitive; they’re the negotiated price of what Opus 5.5 is allowed to do at scale.

The Business Arithmetic

Anthropic has been building toward a two-tier inference economy for most of this year. The documentation framing makes it explicit now. Fable 5.1 is the reasoning model: expensive, maximum-effort by default, for tasks where you need sustained high-effort thinking across a million tokens — complex autonomous agents, frontier mathematical reasoning, the long-horizon tasks where the additional inference compute demonstrably changes the result. Opus 5.5 is the working model: cheaper, medium-effort by default, better than the prior generation at the knowledge work and coding tasks that constitute most enterprise API usage.

The selection decision for builders has sharpened. It used to be: “Do I need Opus 5 or Fable 5.1?” That question was mostly about whether your evals showed improvement and whether the cost was justified. The new question is different: “Does Opus 5.5 clear my bar? If not, which specific thing about Fable 5.1 do I actually need — the higher default effort, the unrestricted capability surface, or reasoning depth that my evals can measure?” Those are answerable questions. They require testing rather than assuming the more expensive model is better.

For Anthropic, the margin logic is clear. A model that captures the majority of enterprise API usage at 40% lower cost per call is a substantially more efficient business than one where the bulk of users are on a $10/$50 model. Fable 5.1’s future depends on the tasks that genuinely require its extended reasoning capacity: long-horizon autonomous work, the cases where the docs now tell you to use it explicitly when “your evals on Claude Opus 5.5 at higher effort still fall short.” That market is real. It is also smaller than “most workloads.”

One pricing detail worth flagging: prompt cache reads on Opus 5.5 cost 5% of the base input price. On Fable 5.1, they’re 2.5%. Heavy cache users running high-cache-hit pipelines on Fable 5.1 have a cost argument that partially erodes the 60% input price advantage. But for workloads without heavy caching — most agentic tasks, most interactive pipelines — the arithmetic still favors Opus 5.5 by a significant margin once the benchmark evidence is taken at face value.

“Start with Claude Opus 5.5 for most workloads” is Anthropic telling its developers that the routing decision has been made for them. Fable 5.1 isn’t deprecated; it remains the right choice for the tasks it was built for. But the default lane changed yesterday. For builders, the honest work is determining whether what you’re running is “most workloads” or the exception — and if you’re routing cybersecurity or biology through Claude, reading the restriction documentation before the constraint surfaces in production.

AI-generated editorial illustration · TemperatureZero · September 23, 2026