# Claude Opus 5.5: benchmarks, pricing, and what's new (Sept 2026)

> Claude Opus 5.5 released Sept 22, 2026: benchmarks vs Opus 5, Fable 5.1 and GPT-6 Astra, pricing ($4 input / $20 output), 1M context, model ID claude-opus-5-5, speed, safety and where to use it.

*Source: https://velstech.net/claude-opus-5-5 · Updated: 2026-09-24 · Category: AI · Tags: LLM, Anthropic, Claude, Benchmarks, AI News, Agents, Coding*

*Markdown version of [Claude Opus 5.5: benchmarks, pricing, and what's new (Sept 2026)](https://velstech.net/claude-opus-5-5). [Read the full guide with interactive tools](https://velstech.net/claude-opus-5-5).*
*Also as Markdown: [Hindi](https://velstech.net/claude-opus-5-5.hi.md) · [Tamil](https://velstech.net/claude-opus-5-5.ta.md).*

---

**Claude Opus 5.5** is Anthropic's new Opus-class flagship, released
**September 22, 2026** — just two months after
Claude Opus 5 (July 24, 2026). It is the first model in the new
**Claude 5.5 family** (Sonnet 5.5 and Haiku 5.5 follow in the coming weeks),
and Anthropic positions it as the enterprise workhorse for long-running coding and knowledge work.

The headline is not just higher scores. Opus 5.5 performs at the level of
[Claude Fable 5.1](https://velstech.net/claude-fable-5-1) on most work while costing
**40% less to run than Opus 5** on typical workloads: lower price per token,
fewer tokens per task, and 60% cheaper cache reads. It is also 30%+ faster at generating output,
writes more clearly, and posts Anthropic's best alignment scores to date.

*Source note:* benchmarks and prices below come from Anthropic's
launch announcement, Claude Platform docs, and early-tester reports cited by Anthropic.
They are vendor-reported, not an independent VelsTech benchmark. Where Anthropic discloses
caveats (safeguard fallbacks, effort settings, standard errors), we keep them.

## The short version: Claude Opus 5.5 in 60 seconds

- What it is: Opus-tier model for agentic coding, computer use, financial research, document work and long investigations. API model ID claude-opus-5-5, 1M-token context, 128K max output, knowledge cutoff June 2026.

- Release date: September 22, 2026, on Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, Claude Code, and GitHub Copilot.

- Price: $4 per million input tokens / $20 per million output tokens — 20% below Opus 5 ($5/$25). Cache reads $0.20/M (60% below Opus 5), 5-min cache writes $5/M, 1-hour writes $8/M, Batch API 50% off. Fast mode ($8/$40) runs up to 2.5x faster.

- Benchmarks: leads Anthropic's reported table on Terminal-Bench 4.0 (66.4%), FrontierCode (54.4%), CursorBench 4.0 (57.8%), GDPval-AA v2.1 (1846 Elo) and OSWorld 2.0 (81.8% partial) — ahead of Opus 5, Fable 5.1, GPT-6 Astra and GPT-5.6 Sol on most of those rows.

- Efficiency: at default (medium) effort it beats Opus 5 at max effort for ~20% of the cost, matches GPT-6 Astra on Terminal-Bench for ~40% of the cost, and beats GPT-5.6 Sol on CursorBench by 11 points for ~one-third the cost.

- Safety: best score to date on Anthropic's automated behavioral audit, ~85% fewer containment-boundary circumvention attempts than Opus 5 / Mythos 5.1, stronger prompt-injection resistance, Fable-5.1-class cyber/bio safeguards, preserved-thinking anti-distillation.

- Who should care: teams running Claude Code, Copilot, or API agents on migrations, audits, financial/legal research and multi-hour jobs. Casual chat users will not feel the difference; heavy agent users will feel it in the bill.

## Claude Opus 5.5 specs at a glance

| Feature | Claude Opus 5.5 | Claude Opus 5 |
| --- | --- | --- |
| Released | Sept 22, 2026 | July 24, 2026 |
| API model ID | claude-opus-5-5 | claude-opus-5 |
| Context window | 1M tokens (default) | 1M tokens |
| Max output | 128K (300K on Batch API with beta header) | 128K (300K batch) |
| Knowledge cutoff | June 2026 | May 2026 |
| Thinking | Adaptive, always on (cannot be disabled) | Adaptive, on by default |
| Default effort | medium (low / medium / high / xhigh / max) | high on API / Code |
| Input / output | Text + images → text | Text + images → text |
| Retirement | Not sooner than Sept 22, 2027 | — |

Platform model IDs: Claude API `claude-opus-5-5`, Bedrock
`anthropic.claude-opus-5-5`, Google Cloud `claude-opus-5-5`,
Microsoft Foundry `claude-opus-5-5`. One request can include up to 600 images or PDF pages.

## Opus 5.5 pricing: what it actually costs

| Per 1M tokens | Opus 5.5 | Opus 5 |
| --- | --- | --- |
| Input | $4 | $5 |
| Output | $20 | $25 |
| Cache reads | $0.20 | $0.50 |
| Cache writes (5 min) | $5 | $6.25 |
| Cache writes (1 hr) | $8 | — |
| Fast mode (up to 2.5x) | $8 / $40 | $10 / $50 |
| Batch API | 50% off input + output |  |

The quiet headline is cache reads: agents re-read the same repo, system prompt, tool definitions
and notes every turn, and cache reads are the majority of coding-agent cost. Cutting them from
$0.50 to $0.20, *plus* using fewer tokens per task, is how Anthropic gets to −40% on typical
workloads and −50% or more on the case studies below. If you estimate agent bills, use our
[AI API cost calculator](https://velstech.net/ai-api-cost-calculator) — remember to count cached input separately.

## Benchmarks: Opus 5.5 vs Opus 5 vs Fable 5.1 vs GPT-6 Astra

All Claude Opus 5.5 rows below use adaptive thinking at max effort unless noted, with production
safeguards enabled. Anthropic warns benchmark margins at this level are a noisy guide to real-world
differences, and that safeguard interventions on cyber/bio tasks were completed by fallback models
(Opus 4.8 / Opus 5), which likely *lowers* the reported Opus 5.5 score there.

| Benchmark (what it tests) | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra |
| --- | --- | --- | --- | --- |
| Terminal-Bench 4.0 (agentic terminal coding) | 66.4% | 55.8% | 52.3% | 57.9% |
| FrontierCode v1.1 Main (would-merge code changes) | 54.4% | 50.3% | 48.0% | 53.3% |
| CursorBench 4.0 (multi-file Cursor sessions) | 57.8% | 51.8% | 46.6% | — |
| GDPval-AA v2.1 (44-occupation knowledge work, Elo) | 1846 | 1735 | 1708 | 1542 |
| AutomationBench (Zapier business workflows) | 40.0% | 31.4% | 26.9% | 41.4% |
| Humanity's Last Exam with tools (multidisciplinary reasoning) | 67.7% | 65.6% | 63.6% | 57.2% |
| Terminal-Bench-Science 0.1 (agentic science) | 58.7% | 52.6% | 29.0% | 64.6% |
| OSWorld 2.0 partial (computer use) | 81.8% | 80.7% | 74.0% | 72.6%* |
| Chartography with tools (chart reading) | 89.0% | 88.4% | 83.4% | — |

* Astra OSWorld figure is the 72.6% reported by OpenAI on its own harness; cross-harness comparisons are directional, not controlled. Terminal-Bench standard error is ±2.6 pts for Opus 5.5; Terminal-Bench-Science ±3.5–5 pts.

Two honest caveats from our reading of [Fable 5.1 vs GPT-6 Astra](https://velstech.net/claude-fable-5-1-vs-gpt-6-astra)
and [our GPT-6 Astra explainer](https://velstech.net/gpt-6-astra): Astra still leads on AutomationBench (41.4 vs 40.0)
and Terminal-Bench-Science (64.6 vs 58.7), and independent indices can rank models differently from vendor tables.
Opus 5.5's clearest sweep is agentic coding + knowledge work + computer use — exactly the enterprise-agent workload it targets.

## Efficiency is the real upgrade

Anthropic's cost-vs-accuracy curves tell the story better than any single score: Opus 5.5 at
**default (medium) effort** already beats Opus 5 at max effort for about a fifth of the cost
on Terminal-Bench, beats GPT-6 Astra's top FrontierCode score for ~20% of the per-task cost, and matches
Astra on Terminal-Bench for ~40% of the cost.

- 200,000-line audit + fix: under 3 hours on Opus 5.5 vs 20+ hours and 2.5x tokens on Opus 5 (early tester).

- 680,000-line migration: under a day on Opus 5.5 vs weeks for an engineering team.

- HAProxy C → Rust rewrite: both Opus 5.5 and Fable 5.1 passed nearly all regression tests, but Opus 5.5 finished in 9.5 vs 12 hours at 51% lower cost.

- Web-app load-time cuts: 39/40 pages improved without behavior changes (Opus 5 made smaller gains that altered behavior).

- Merger model + deck: same conclusion as Opus 5, but more thorough, easier to read, minor errors fixed, 63 vs 93 minutes at 50% lower cost.

Partner reports rhyme: GitHub measured among the fewest tokens/steps in Copilot CLI and VS Code
(more terminal tasks solved in under half the steps vs Opus 5); Lovable reports one-third to one-half
fewer steps; Kiro reports ~40% fewer calls and half the tokens; Optiver reports half the turns/time/output
tokens for equal quality (−40 to −50% cost); Box reports one-third the tokens at 40% lower verbosity.

## Coding and computer use

Opus 5.5 is built for sprawling jobs — codebase-wide migrations, audits, multi-repo refactors —
where the agent must gather context once, delegate to subagents, verify its own work, and keep going
for hours. Column reports better subagent delegation and self-verification loops; Clio left it running
unattended overnight for 18+ hours across six repos; Stripe ran one session directing a dozen more
through a 40-PR rebase with all 40 passing CI next afternoon.

On computer use it scores 81.8% partial on OSWorld 2.0, and Anthropic ships a hardened agent stack for
unattended runs: a classifier screening every action, an open-source sandbox security teams can audit,
and code review that catches vulnerabilities before merge. Gray Swan's prompt-injection tests tie it with
Fable 5.1 for the lowest injection success rate of any model tested.

## Knowledge work: research, finance, legal

On GDPval-AA v2.1 Opus 5.5 scores 1846 Elo vs 1735 (Fable 5.1) and 1708 (Opus 5), and at medium effort
beats GPT-6 Astra at max effort for ~one-fifth the per-task cost. In Anthropic's hard-citation test
(quarterly-performance report where the earnings release was hard to find, any invented figure/quote = fail),
16 of 18 Opus 5.5 reports cleared the bar; Fable 5.1 and Opus 5 cleared zero.

Early testers: Walleye Capital says it largely solved their quant eval on the *lowest* setting and
caught an off-by-one in their own instructions no prior model caught; Deloitte reports 72% of known bugs
caught in code review at lowest effort vs 56% for Opus 5 at high effort; Hebbia reports 86.6% rubric coverage
vs 60.3% with best-ever citation recall; LexisNexis and Thomson Reuters report stronger citations and expert-eval wins.

## Communication: it writes like a colleague now

The most quoted tester line — "it writes the way I do" — reflects deliberate work: most important
information first, less jargon, fewer idiosyncratic phrases, better adherence to your writing rules.
In Anthropic's side-by-sides, Opus 5.5 quantifies the impact ($1.50 vs $9.92), states ownership and next
steps in thread summaries, and explains *why* it rejected your design (blocked lines need
between-masks, not fixed masks) instead of narrating tensor internals. Testers say long sessions are
easier to follow and check — which Anthropic frames as a safety benefit too.

## Safety and safeguards: strongest audit scores, stricter defaults

- Alignment: best scores to date on Anthropic's ~2,000-scenario automated behavioral audit; stronger on honesty; ~85% fewer containment-boundary circumvention attempts than Opus 5 / Mythos 5.1 (every remaining attempt low-severity and self-reported); more resistant to prompt injection.

- Cyber: Fable-5.1-class safeguards. Routine bug-fix stays in-product; most offensive cyber tasks re-route to Opus 4.8. Verified defenders get expanded Cyber Verification Program tiers (including Mythos access). Claude Security already offers Mythos 5.1.

- Biology: matches/beats Mythos 5.1 on several bio evaluations, so same safeguards as Fable 5.1. Vetted labs/startups/pharma can apply to the Life Sciences Verification Program.

- Distillation: preserved-thinking anti-distillation (API users cannot edit prior context to extract reasoning), for API accounts created on/after Aug 31, 2026.

- Compliance: zero data retention available; EU AI Act watermarking like Fable 5.1; thinking cannot be switched off.

Anthropic pairs this with a caveat worth keeping: Opus 5.5 often suspects it is being evaluated, which
limits how confidently lab behavior predicts open-world behavior, and building evaluations that catch
every failure pre-deployment remains unsolved. Treat the audit wins as directional, not a guarantee —
especially if you run agents unattended for hours.

## Availability: where to use Claude Opus 5.5 today

- Claude API / Platform: claude-opus-5-5 — Claude Code, Cowork, Claude in Chrome, Microsoft 365 integration.

- Clouds: Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, Claude Platform on AWS.

- GitHub Copilot: Pro+, Max, Business and Enterprise via the model picker (VS Code, Visual Studio, Copilot CLI, coding agent, JetBrains, Xcode, Eclipse, Mobile) — gradual rollout, admin-managed by model policy.

- Subscriptions: higher five-hour limits on Pro / Max / Team / seat-based Enterprise, plus a save-for-later rate-limit reset.

- Coming weeks: Claude Sonnet 5.5 and Claude Haiku 5.5 with the same efficiency/safety direction.

## Opus 5.5 vs Opus 5: should you switch?

If you already use Opus 5, the migration is a model-ID change plus retesting effort/thinking behavior —
but the economics favor switching fast for agent workloads: −20% sticker price, −60% cache reads, fewer
tokens per task, 30%+ faster output. The only reason to stay is a frozen production prompt that depends
on Opus 5 quirks (including its more verbose style) — pin it, test side-by-side at medium vs high effort,
and watch safeguard re-routes on cyber/bio tasks.

If you use Fable 5.1 for frontier reasoning, keep it for now: Anthropic says the real-world gap between
Opus 5.5 and Fable 5.1 is narrower than benchmarks suggest, and Fable remains the top-tier brain at
$10/$50. If you use Sonnet for volume, wait for Sonnet 5.5 — Opus 5.5 is the quality/efficiency play,
not the cheapest-seat play.

## What it means for local-AI users

Nothing to download: Opus 5.5 is cloud-only, like
[Fable 5.1](https://velstech.net/claude-fable-5-1) and [GPT-6 Astra](https://velstech.net/gpt-6-astra).
Your RX 6800M, RTX card or Mac still runs open weights locally — private, offline, predictable cost —
while Opus pulls further ahead on multi-hour autonomy no 27–35B local model can match.
See [best GPU for local LLMs](https://velstech.net/best-gpu-for-local-llm),
[how much VRAM you need](https://velstech.net/how-much-vram-for-llm) and
[local vs cloud AI](https://velstech.net/local-vs-cloud-ai) for the tradeoff math.

Borrow the workflow, not the weights: durable notes across context windows, checkpointed long jobs,
test-after-every-change, evidence-over-confidence reports, and subagent delegation with self-checks.
Those patterns lift a 7B local agent too — and when frontier efficiency tricks cascade into open
weights (as they always do), our [ROCm vs Vulkan guide](https://velstech.net/rocm-vulkan-amd-guide) and
[benchmark database](https://velstech.net/benchmarks/index) will show what fits your VRAM.

## Bottom line

Claude Opus 5.5 is the rare flagship that is both more capable *and* meaningfully cheaper:
Fable-class coding, research and computer use at Opus prices cut nearly in half in practice, with
clearer writing and the strongest alignment audit Anthropic has reported. It does not win every row
(Astra takes AutomationBench and Terminal-Bench-Science in Anthropic's own table), and vendor tables
are not independent tests — but for teams delegating real multi-hour software and knowledge work,
it is the new default to beat. Sonnet 5.5 / Haiku 5.5 will decide how far down the stack the 5.5
efficiency story goes.

## Sources

- Anthropic: Introducing Claude Opus 5.5 (Sept 22, 2026)

- Claude Platform Docs: Claude Opus 5.5 overview (model IDs, pricing, limits)

- Claude Opus 5.5 System Card (PDF)

- Our Claude Fable 5.1 explainer

- Our GPT-6 Astra explainer

- Fable 5.1 vs GPT-6 Astra comparison

## FAQ

**What is Claude Opus 5.5?**

Anthropic's Opus-class flagship released Sept 22, 2026 — the first Claude 5.5 model, built for long-running agentic coding, computer use and knowledge work at Fable-5.1-level quality and ~40% lower cost than Opus 5.

**When was Claude Opus 5.5 released?**

September 22, 2026, two months after Claude Opus 5 (July 24, 2026). Sonnet 5.5 and Haiku 5.5 follow in the coming weeks.

**How much does Claude Opus 5.5 cost?**

$4 per million input tokens and $20 per million output tokens, with $0.20 cache reads, $5 5-min cache writes, $8 1-hr writes, 50% off on Batch API, and $8/$40 Fast mode. Typical workloads cost ~40% less than Opus 5.

**What is the Claude Opus 5.5 model ID and context window?**

API ID `claude-opus-5-5` (Bedrock: `anthropic.claude-opus-5-5`), 1M-token context, 128K max output (300K batch), June 2026 cutoff, adaptive thinking always on with low/medium/high/xhigh/max effort.

**Is Claude Opus 5.5 better than Opus 5, Fable 5.1 or GPT-6 Astra?**

On Anthropic's reported table it leads Opus 5 everywhere and Fable 5.1 on most coding/knowledge rows, and beats GPT-6 Astra on Terminal-Bench, FrontierCode and GDPval while trailing on AutomationBench and Terminal-Bench-Science. Real-world gaps are narrower than benchmarks suggest.

**Where is Claude Opus 5.5 available?**

Claude API/Platform, Claude Code, Amazon Bedrock, Google Cloud, Microsoft Foundry, and GitHub Copilot (Pro+, Max, Business, Enterprise) with gradual rollout. Higher five-hour limits on Pro/Max/Team/Enterprise.

**Can I run Claude Opus 5.5 locally?**

No. It is closed and cloud-only. Use the API/clouds/Copilot, or run open-weight models locally for privacy/offline work — see our local-AI guides.

## FAQ

**What is Claude Opus 5.5?**

Anthropic's Opus-class flagship released Sept 22, 2026 — the first Claude 5.5 model, built for long-running agentic coding, computer use and knowledge work at Fable-5.1-level quality and about 40% lower cost than Opus 5.

**How much does Claude Opus 5.5 cost?**

$4 per million input tokens and $20 per million output tokens, with $0.20 cache reads, $5 5-min cache writes, 50% off on Batch API, and $8/$40 Fast mode. Typical workloads cost about 40% less than Opus 5.

**What is the Claude Opus 5.5 model ID and context window?**

API ID claude-opus-5-5 (Bedrock: anthropic.claude-opus-5-5), 1M-token context, 128K max output (300K batch), June 2026 cutoff, adaptive thinking always on with low/medium/high/xhigh/max effort.

**Is Claude Opus 5.5 better than Opus 5 or GPT-6 Astra?**

On Anthropic's reported table it leads Opus 5 everywhere and beats GPT-6 Astra on Terminal-Bench 4.0, FrontierCode and GDPval, while trailing on AutomationBench and Terminal-Bench-Science. Real-world gaps are narrower than benchmarks suggest.

**Where is Claude Opus 5.5 available?**

Claude API/Platform, Claude Code, Amazon Bedrock, Google Cloud, Microsoft Foundry, and GitHub Copilot (Pro+, Max, Business, Enterprise). Higher five-hour limits on Pro/Max/Team/Enterprise.

**Can I run Claude Opus 5.5 locally?**

No. It is closed and cloud-only. Use the API, clouds or Copilot — or run open-weight models locally for privacy and offline work.

---

*VelsTech – technology explained for everyone. Original: https://velstech.net/claude-opus-5-5*