Claude Opus 5.5 Hits Fable-Level Work at 40% Less Than Opus 5 — What Developers Get

Anthropic released Claude Opus 5.5 on September 22, 2026 — the first model in its Claude 5.5 family. The pitch is blunt: Fable-level work for most tasks, at roughly 40% less cost than Opus 5 on typical workloads.

List prices confirm the cut. Input and output tokens are $4 and $20 per million, down from Opus 5’s $5 and $25. Cache reads — the line item that dominates long agent loops — drop from $0.50 to $0.20. The API ID is claude-opus-5-5, and it is live on Anthropic’s API, Amazon Bedrock, Google Cloud, and Microsoft Foundry.

The same afternoon, OpenAI shipped GPT-6 Sol and Luna. TechCrunch noted Opus 5.5 landed about 90 minutes earlier. Both labs are competing on cost per finished task, not raw scoreboard bragging. Treat Anthropic’s benchmark table as vendor-reported until independent leaderboards catch up.

What Anthropic Actually Announced

In the launch post, Anthropic says Opus 5.5 “performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.” It is the company’s first release since CEO Dario Amodei called for pacing the frontier.

Anthropic also says the model was tested before release by external evaluators, including Frontier Design and METR, and that Opus 5.5 posted the strongest scores yet on Anthropic’s automated behavioral audit. Sonnet 5.5 and Haiku 5.5 are promised “in the coming weeks,” with many of the same efficiency and safety changes.

Platform docs list the model as active, released September 22, 2026, with retirement “not sooner than September 22, 2027.” Knowledge cutoff is June 2026. Context window is 1 million tokens; max output is 128K on the synchronous Messages API (up to 300K on Message Batches with a beta header).

The Rate Card, Straight From Anthropic

These are standard Claude API rates from Anthropic’s pricing page and the Opus 5.5 model page, in USD per 1 million tokens:

Token list price is a 20% cut on input and output versus Opus 5. Cache reads are 60% cheaper — Anthropic says those reads make up most of the cost on agentic and coding work. Cache writes are $5 (5-minute) and $8 (1-hour) per million tokens. Batch API is half price: $2 / $10.

Fast mode is a separate research preview on the Claude API only (not Bedrock, Google Cloud, or Foundry). Per the pricing docs, Opus 5.5 Fast is $8 / $40 — double standard rates — for up to 2.5× output speed. Anthropic’s launch post cites the same Fast rates for Claude Code and the Claude Platform.

Anthropic’s “40% less on typical workloads” claim is not the same as the 20% list-price cut. It bundles cheaper tokens, fewer tokens per task, and faster completion. That is a vendor workload estimate. Run your own traffic before you budget against it.

Four Breaking API Changes

Anthropic’s What’s new page is clear: swapping the model ID is not enough if you are coming from Opus 5.

1. Thinking cannot be disabled. thinking: {"type": "disabled"} or a manual budget_tokens budget returns a 400. Use adaptive thinking and control depth with output_config.effort.
2. Forced tool use is gone. tool_choice of any or a named tool returns a 400. Stick to auto or none; use strict tools or structured outputs for schema-valid JSON.
3. Thinking blocks are bound to the model and conversation. Opus 5.5 reads thinking from earlier Opus/Sonnet/Haiku models, but not from Fable or Mythos. Editing the system prompt, tools, or earlier messages after a thinking block can error on accounts created on or after August 31, 2026 (preserved thinking / anti-distillation).
4. Old computer-use tool rejected on Claude API and Google Cloud. computer_20251124 returns a 400; move to computer_toolset_20260801. Bedrock still accepts the earlier tool on Opus 5.5.

Behavior shifts that do not throw errors matter just as much. Default effort is now medium (Opus 5 defaulted to high). Text between tool calls arrives in thinking blocks, and at the default display: "omitted" streaming UIs go quiet mid-agent-loop until you opt in. More safeguard categories can refuse biology and reasoning-extraction requests with stop_reason: "refusal".

What the Benchmarks Claim (Vendor Numbers)

Anthropic’s launch table is Anthropic’s own runs, usually with adaptive thinking at max effort unless noted. Competitor scores for GPT-6 Astra and GPT-5.6 Sol are “as reported by OpenAI.” Label all of this vendor-reported.

Anthropic’s footnotes matter. Production safeguards were on during evals; when they intervened, cyber tasks fell back to Opus 4.8 and biology / frontier-LLM-dev tasks to Opus 5, which likely lowered some scores. AutomationBench runs treated safeguard interventions as failures. Anthropic itself says the real-world gap to Fable 5.1 is “narrower than these scores suggest.”

The cost-per-task charts are the sharper claim: at default effort on FrontierCode, Anthropic says Opus 5.5 beats GPT-6 Astra at roughly one-fifth the cost per task; on Terminal-Bench 4.0 it matches Astra at about 40% of the cost. Those are Anthropic’s cost models on Anthropic’s harness. Independent replication is still the test.

Safeguards, Alignment, and the Pacing Pitch

Because Anthropic rates Opus 5.5 as comparable to Mythos 5.1 in biology and cybersecurity, it ships with Fable-class safeguards. Routine bug-finding stays available; most cybersecurity tasks re-route to Opus 4.8. Biology research that hits the classifier needs the Life Sciences Verification Program. A Cyber Verification Program expansion is promised for vetted defenders.

Preserved thinking — the anti-distillation control introduced with Fable 5.1 — applies to Opus 5.5 for API accounts created on or after August 31, 2026. Zero data retention remains available. Thinking mode can no longer be switched off.

On Anthropic’s automated behavioral audit (~2,000 scenarios), the company says Opus 5.5 improved on behaviors linked to recent agent incidents — motivated reasoning, sandbox-escape attempts, and harmful actions after concluding it was in a simulation. In a containment-boundary eval, Anthropic says it attempted to circumvent boundaries about 85% less often than Opus 5 or Mythos 5.1. That is Anthropic measuring Anthropic; the system card is the full package.

The political framing is hard to miss. The launch post spends a long section on pacing the frontier while still shipping a frontier coding agent the same day OpenAI cut Sol and Luna in half. For the Amodei policy thread this week, see our coverage of the White House dinner confirmation and Australia’s Senate AI inquiry invitations.

What It Means for Indian Developers

Opus 5.5 at $4 / $20 lands between Sonnet 5 ($2 / $10) and Fable 5.1 ($10 / $50), and sits next to OpenAI’s Sol at $2 / $10. For Bangalore and Hyderabad teams running overnight coding agents, the cache-read cut is the practical win: long contexts with heavy reuse get cheaper faster than the headline 20% list cut suggests.

A sensible stack for cost-sensitive Indian startups: Sonnet 5 or GPT-6 Luna for extraction and routing, Opus 5.5 or Sol for multi-hour coding agents, Fable or Astra only when failure is expensive. Compare those list prices with open-weight options we have tracked, including Qwen 3.8-27B and DeepSeek V4 Pro pricing.

Budget for migration. Forced tool use, disabled thinking, and old computer-use tool IDs will break some Opus 5 integrations on day one. Re-sweep effort: default medium is not the old high. Our earlier piece on Claude Opus 5 covers the bar this release is priced against.

What’s Confirmed vs. Still Open

Frequently Asked Questions

How much does Claude Opus 5.5 cost?

On the Claude API: $4 per million input tokens and $20 per million output tokens. Cache reads are $0.20; 5-minute cache writes are $5; 1-hour writes are $8. Batch is $2 / $10. Fast mode (Claude API only) is $8 / $40.

What is the Claude Opus 5.5 model ID?

claude-opus-5-5 on the Claude API, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. On Amazon Bedrock it is anthropic.claude-opus-5-5.

Is Claude Opus 5.5 better than Claude Fable 5.1?

Anthropic says it performs at Fable 5.1 level on most work and leads several of Anthropic’s own coding and knowledge-work benches, while costing far less per token. Anthropic also says the real-world gap to Fable is narrower than the scoreboard. Independent benches will settle the argument.

Will my Claude Opus 5 code keep working?

Not if you disable thinking, force tool choice, or still call computer_20251124 on the Claude API or Google Cloud. Update the model ID, switch to adaptive thinking plus effort, use tool_choice: auto, and migrate computer use to the 2026 toolset. See Anthropic’s migration guide.

When do Sonnet 5.5 and Haiku 5.5 ship?

Anthropic only says “in the coming weeks.” No dates or prices have been published for those two as of this writing.