WayToClawEarn
High impactAnthropic

Claude Opus 5.5 launched: 40% cheaper, new coding SOTA — which model to pick?

Anthropic launched Claude Opus 5.5 on September 22, matching Fable 5.1 on most benchmarks at 40% of the per-token cost. Input $4/M, output $20/M — 20% cheaper than Opus 5, 30%+ faster output. SWE-bench Pro hits 89.9 (new SOTA), Terminal-Bench 4.0 hits 66.4.

Edisen Lu · WayToClawEarnVia AnthropicPublished Sep 28, 2026

Reviewed from public sources · AI-assisted drafting under editorial oversight. How we work · Original source

Public-source compilation

Synthesized from public posts/docs. Prefer the original source for primary claims.

How we review content · Primary source · Anthropic

TL;DR

If you searched for "Claude Opus 5.5 pricing" or "Opus 5.5 vs Fable 5.1": Anthropic launched Claude Opus 5.5 on September 22, matching or beating Fable 5.1 on most benchmarks at 40% of the per-token cost. Input is $4/M tokens, output is $20/M — 20% cheaper than Opus 5, with 30%+ faster output generation.

Pricing Comparison (per million tokens, USD)

CategoryOpus 5.5Opus 5Fable 5.1
Input$4$5$10
Output$20$25$50
Cache write (5min)$5$6.25$12.50
Cache read$0.20$0.50$0.25
Batch input/output$2/$10$2.50/$12.50$5/$25
Fast mode input/output$8/$40$10/$50Not offered

Key advantage: cache reads at 5% of input price (typical is 10%), a major saving for agent loops that re-read stable prefixes. The full 1M context bills at standard rate with no long-context premium.

Key Benchmark Results

BenchmarkOpus 5.5Opus 5Fable 5.1
SWE-bench Pro89.979.281.2
SWE-bench Multilingual93.989.589.1
Terminal-Bench 4.066.452.355.8
FrontierCode v1.154.448.050.3
Humanity's Last Exam (no tools)64.456.660.9
OSWorld 2.0 (partial/strict)81.8/48.774.0/37.280.7/42.8
CursorBench 4.057.846.651.8

Coding is the headline: SWE-bench Pro jumps +10.7 pts over Opus 5; Terminal-Bench 4.0 jumps +14.1 pts (new state of the art).

Model Specs

ParameterValue
Context window1,000,000 tokens
Max output (Messages API)128,000 tokens
Max output (Batch API)300,000 tokens
Knowledge cutoffJune 2026
Model IDclaude-opus-5-5
PlatformsClaude API, Bedrock, Vertex AI, Foundry
Earliest retirementSeptember 22, 2027

Note: Models since Opus 4.7 use a newer tokenizer producing ~30% more tokens for the same text. A 1M window holds ~555K English words (vs ~750K on older models).

Speed Improvements

  • Output generation 30%+ faster than Opus 5
  • Fast mode up to 2.5x faster (research preview, first-party API only)
  • 40% less cost on typical workloads (Anthropic's claim, holds at medium/high effort)
  • Customer reports: Optiver matched Opus 5 quality "in about half the turns, time and output tokens"; Kiro saw ~40% fewer calls and half the tokens

Developer Selection Guide

Use caseRecommendedWhy
Coding, multi-step agent tasksOpus 5.5SWE-bench Pro new SOTA at 40% of Fable 5.1 cost
Hardest long-horizon reasoningFable 5.1OfficeQA / OfficeQA Pro still leads
High-frequency agent loops (cache-heavy)Opus 5.5Cache reads at 5% of input price
Need Fast modeOpus 5.5Fable 5.1 does not offer Fast

Migration Notes

  1. Tokenizer change: New tokenizer since Opus 4.7 produces ~30% more tokens for the same text — recalculate budget estimates
  2. Model ID: Use claude-opus-5-5 directly, no date-suffixed snapshot ID
  3. Batch API large output: Add output-300k-2026-03-24 header to unlock 300K output
  4. Fast mode: Requires speed: "fast" and fast-mode-2026-02-01 beta header, first-party API only
  5. US-only inference: Setting inference_geo: "us" adds a 10% surcharge

Next Steps

  1. Test claude-opus-5-5 in your Claude Code or API projects, compare cost/quality vs your current model
  2. If using agent loops, enable prompt caching and evaluate cache-read savings
  3. See Anthropic docs for full API reference

Related: GPT-6 Sol and Luna launched the same day with 50% API price cuts — read the coverage.

anthropicclaudeopus-5-5apipricingcodingllm

View source →

Educational reference only: cases summarize public sources and may use AI-assisted drafting under editorial review. Not financial advice; outcomes are not guaranteed.