Claude Opus 5.5: Fable-Class Results, Lower Cost

Claude Opus 5.5: Fable-Class Results at a Fraction of the Cost

On September 22, 2026, Anthropic shipped Claude Opus 5.5, the first model in a new Claude 5.5 family. The positioning is unusually blunt for a frontier lab: perform at Claude Fable 5.1's level on most work, and cost roughly 40% less to run than Opus 5 at default settings — not fudge the margins, just be clear about it.

The numbers back that story. Opus 5.5 lists at $4 per million input tokens and $20 per million output tokens, down from Opus 5's $5/$25. Cache reads drop to $0.20 per million (from $0.50, a 60% cut), output is generated more than 30% faster, and Anthropic's arithmetic says a typical workload nets out at about 40% less total cost because the model also burns fewer tokens per task.

It's also the first frontier release since CEO Dario Amodei called for pacing the industry. And it comes with a wrinkle agent builders should not ignore: safety classifiers can silently hand your request to an older model mid-workflow.

What Opus 5.5 actually is

Opus 5.5 keeps the 1M-token context window and 128K max output of its predecessor, with a June 2026 knowledge cutoff and the model ID claude-opus-5-5 on the Claude API plus Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. Anthropic won't say how big it is, only that it's smaller and cheaper to serve than Fable 5.1. Sonnet 5.5 and Haiku 5.5 are due "in the coming weeks" with the same efficiency play.

The pricing is the headline:

Fable 5.1's base token rates are 60% higher than Opus 5.5's on every line. The one line where Fable still looks proportionally better is the cache discount ratio — $0.25 is only 2.5% of its $10 input rate, while Opus 5.5's $0.20 is 5% of $4. For a pipeline dominated by cache hits, that math still favors Fable on a per-cached-token basis. Worth knowing before you mass-migrate.

The benchmark table, with Anthropic's own asterisk

Anthropic's published numbers put Opus 5.5 ahead on most rows it chose to show:

All of it is self-reported, and the footnote is honest about conditions: production safeguards were on, and when a safety classifier fired, cybersecurity tasks were completed by Opus 4.8 and biology-plus-frontier-LLM tasks by Opus 5 — which Anthropic says likely reduces Opus 5.5's score on those benches. GPT-6 Astra still out-scores it on AutomationBench and Terminal-Bench-Science, so "best at everything" it is not.

Then there's the caveat Anthropic itself wrote, which is rarer than the scores: "at these levels of capability we've found that benchmark margins have become a less reliable guide to real-world differences. In our own use, the gap between Opus 5.5 and Claude Fable 5.1 is narrower than these scores suggest." An independent re-run by Artificial Analysis measured Opus 5.5 at 59.6% on Terminal-Bench 4.0 against Anthropic's 66.4%. Treat the table as directional, not gospel.

The efficiency claim is where Anthropic wants the story to land. At default effort, it says Opus 5.5 beats GPT-6 Astra on FrontierCode at roughly 20% of the per-task cost, matches Astra on Terminal-Bench 4.0 at about 40% of the cost, and beats GPT-5.6 Sol by 11 points on CursorBench at about a third of the cost.

Migration notes for agent builders

Adaptive thinking is now always on — you cannot turn it off on Opus 5.5. You control depth with the effort parameter, and the API default is medium, not high. For code that already sets thinking, the docs list four migration gotchas: thinking can't be disabled, forced tool use returns an error, thinking blocks are tied to the model and conversation, and the computer_20251124 computer-use tool is rejected. The first three also apply to Fable 5.1, if you migrated there first.

A quick sketch of the shape of the change:

// Opus 5: thinking was opt-in.
await anthropic.messages.create({
  model: "claude-opus-5",
  max_tokens: 8192,
  thinking: { type: "enabled", budget_tokens: 12000 },
  messages: [{ role: "user", content: "review this diff" }],
});

// Opus 5.5: thinking is always on; set effort, default is "medium".
await anthropic.messages.create({
  model: "claude-opus-5-5",
  max_tokens: 8192,
  effort: "medium",
  messages: [{ role: "user", content: "review this diff" }],
});

The deeper gotcha is invisible in these calls. The New Stack flagged it clearly: Opus 5.5 ships with the same class of safety classifiers Anthropic already runs on Fable 5.1, and when one fires, the request is transparently rerouted to an older model — most cybersecurity flags to Opus 4.8, biology and frontier-LLM-development flags to Opus 5. The user cannot see it happen. If you build evals or agent workflows on the assumption that every request hits the same capability, a rerouted step can change downstream behavior mid-run without a trace in your logs. That's an architectural note, not a reason to panic — but it absolutely belongs on the migration checklist.

The HAProxy story and the efficiency case

Anthropic's internal head-to-head is the nice one. It asked Opus 5.5 and Fable 5.1 to translate HAProxy, the load balancer, from C into Rust. Both rewrites passed nearly all of HAProxy's own regression tests, but Opus 5.5 finished in 9.5 hours versus 12, and cost 51% less. That's the entire pitch in one paragraph: same class of result, less money, less wall-clock.

A few customer anecdotes point the same direction, with the usual caveat that they are vendor-shared, not independent benchmarks:

• Box reported Opus 5.5 used about a third as many tokens as Opus 5 in its evaluations, with answers 40% less verbose at the same accuracy.
• GitHub tested it inside Copilot CLI and VS Code and found it completed more terminal tasks than Opus 5 in less than half the steps.
• Deloitte said Opus 5.5's lowest-effort setting caught 72% of known bugs in code reviews, against 56% for Opus 5 at high effort, with fewer false alarms.

And you can feel the efficiency story in your own wallet with a napkin calculation, since the whole cost claim is just price times tokens:

const opus5  = { input: 5, output: 25, cache: 0.5 };
const opus55 = { input: 4, output: 20, cache: 0.2 };

function cost(r, fresh, cached, out) {
  return r.input * (fresh / 1e6) + r.cache * (cached / 1e6) + r.output * (out / 1e6);
}

const run = cost(opus55, 40_000, 760_000, 40_000); // ≈ $1.11
const old = cost(opus5,  40_000, 760_000, 40_000); // ≈ $1.58
console.log(`Opus 5.5 $${run.toFixed(2)} vs Opus 5 $${old.toFixed(2)}`);
// → about 30% cheaper before the fewer-tokens effect (Anthropic's 40% figure adds that).

Should you migrate?

Anthropic's own guidance is refreshingly specific: use Opus 5.5 for most work, keep Fable 5.1 reserved for the hard, open-ended tasks that need extra reasoning, and leave Opus 5 running only for integrations already pinned to it. That matches the other lesson from this launch season — the harness economist's warning on Strands that token price is only half of agent cost, and effort settings matter more than the base rate.

The benchmark gap between these frontier models is now smaller than the noise around it, and Anthropic basically says so on the tin. The practical move is unchanged: pick the effort level you'll actually run in production, re-measure your workload's cost and wall-clock on both models, and let your own evals — not the launch table — decide. For most teams, Opus 5.5's math is going to win. For the genuinely stubborn tasks, Fable 5.1 is still there, and it costs what it costs.

Sources

• Anthropic: Introducing Claude Opus 5.5
• Anthropic: Claude Opus 5.5 System Card
• Anthropic: Claude Opus 5.5 — Platform documentation
• The New Stack: Anthropic releases Opus 5.5 and cuts pricing by 20%. Your agent calls might secretly get routed to an older model.
• Anthropic: Introducing Claude Fable 5.1 and Claude Mythos 5.1
• TechCrunch: Anthropic releases Opus 5.5 with lower prices and Fable-level performance