Claude Opus 5.5: Fable 5.1 Performance at 40% Less Cost, the First Release After the Pacing Call
On September 22, 2026, Anthropic released Claude Opus 5.5 — the first model in the Claude 5.5 family, delivering Fable 5.1-level performance at 40% lower cost than Opus 5, with best-in-class safety scores, preserved thinking anti-distillation safeguards, and the first release since CEO Dario Amodei's call for frontier pacing.
Claude Opus 5.5: Fable 5.1 Performance at 40% Less Cost, the First Release After the Pacing Call
Executive Summary
On September 22, 2026, Anthropic released Claude Opus 5.5, the first model in its new Claude 5.5 family. This release represents a strategic recalibration of Anthropic's product lineup: Opus 5.5 delivers performance comparable to the higher-tier Claude Fable 5.1 on most work while costing 40% less to run than its predecessor, Claude Opus 5.
Opus 5.5 is Anthropic's first release since CEO Dario Amodei's September 12 call for a global AI slowdown ("We Must Pace the Frontier"). It was tested before release by external evaluators including Frontier Design and METR, and achieved the best scores of any model to date on Anthropic's automated behavioral audit — the most comprehensive alignment test the company runs. The model is significantly less likely to take hard-to-reverse actions, more resistant to prompt injection, and deploys with safeguards similar to those on Fable 5.1, including the preserved thinking anti-distillation mechanism.
Pricing is set at $4/M input and $20/M output — 20% less than Opus 5 — with cache reads at $0.20/M (60% less than Opus 5). On benchmarks, Opus 5.5 leads in agentic coding (66.4% on Terminal-Bench 4.0), knowledge work (1846 Elo on GDPval-AA v2.1), and computer use (81.8% on OSWorld 2.0). Early testers reported dramatic efficiency gains: a 680,000-line code migration completed in less than a day, a 200,000-line codebase audit in under three hours (vs. 20+ hours for Opus 5), and a HAProxy C-to-Rust translation at 51% lower cost than Fable 5.1.
This article extends the September 2026 frontier model survey alongside Claude Fable 5 1 Mythos 5 1 Scientific Frontier Agentic Coding Enterprise Frontier Safeguards 2026 09 17 (which covered the Fable 5.1 and Mythos 5.1 launch on September 1) and Ai News Week 2026 09 14 2026 09 21 (which covered Amodei's pacing call and the broader September landscape).
1. The Pacing Context: First Release After the Slowdown Call
1.1 Amodei's September 12 Essay
On September 12, 2026, Anthropic CEO Dario Amodei published a 3,800-word essay titled "We Must Pace the Frontier," calling for a global AI slowdown. His three-step plan:
- Invite independent reviewers into Anthropic to evaluate models before release
- Coordinate safety checkpoints with other labs
- Work with governments on regulatory frameworks
The call was backed by OpenAI CEO Sam Altman and xAI founder Elon Musk, creating unusual alignment among competing CEOs. Pushback was swift: industry insiders framed the safety push as "regulatory capture" designed to protect margins from open-weight competition.
1.2 Opus 5.5 as a Pacing Release
Opus 5.5 is explicitly positioned as Anthropic's response to its own pacing call:
"Claude Opus 5.5 is our first release since we called for pacing the frontier. It was tested before release by external evaluators, including Frontier Design and METR."
This distinguishes Opus 5.5 from previous releases in several ways:
| Aspect | Previous Releases | Opus 5.5 |
|---|---|---|
| External evaluation | Internal testing only | Frontier Design + METR pre-release review |
| Behavioral audit | Standard alignment suite | Best scores of any model tested |
| Safety margin | Standard | Temporarily wider margin against jailbreaks |
| Safeguards | Standard for tier | Fable 5.1-level safeguards |
| Anti-distillation | Fable 5.1 only | Extended to Opus 5.5 |
1.3 The Strategic Signal
The Opus 5.5 release sends a clear signal: Anthropic is prioritizing efficiency and safety over raw capability escalation. Rather than releasing a more capable model, they released a model that matches the intelligence of their best (Fable 5.1) at a fraction of the cost, with stronger safety guarantees. This aligns with the pacing narrative — demonstrating that progress doesn't require constantly pushing the capability frontier.
2. Performance Benchmarks: Leading Across Categories
2.1 Agentic Coding
Opus 5.5 leads all models on key agentic coding benchmarks:
| Benchmark | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 55.8% | 52.3% | 57.9% | 37.3% |
| FrontierCode v1.1 | 54.4% | 50.3% | 48.0% | 53.3% | 47.5% |
| CursorBench 4.0 | 57.8% | 51.8% | 46.6% | — | 41.7% |
On Terminal-Bench 4.0, Opus 5.5 achieves 66.4% at default effort — beating Opus 5 at max effort (52.3%) for approximately one-fifth of the cost. It matches GPT-6 Astra's best score for about 40% of the cost.
2.2 Knowledge Work
| Benchmark | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|---|---|---|
| GDPval-AA v2.1 | 1846 Elo | 1735 | 1708 | 1542 | 1588 |
| AutomationBench | 40.0% | 31.4% | 26.9% | 41.4% | 28.8% |
On GDPval-AA v2.1 — a test of real-world work across 44 occupations — Opus 5.5 scores 1846 Elo, ahead of Fable 5.1 (1735) and Opus 5 (1708). At default effort (medium), it beats GPT-6 Astra at max effort for about one-fifth of the cost per task.
2.3 Scientific Research & Computer Use
| Benchmark | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Terminal-Bench-Science 0.1 | 58.7% | 52.6% | 29.0% | 64.6% | 22.4% |
| OSWorld 2.0 | 81.8% (partial) | 80.7% (partial) | 74.0% (partial) | — | — |
| Humanity's Last Exam | 67.7% (with tools) | 65.6% (with tools) | 63.6% (with tools) | 57.2% (with tools) | — |
| Chartography | 89.0% (with tools) | 88.4% (with tools) | 83.4% (with tools) | — | — |
On Terminal-Bench-Science 0.1, Opus 5.5 more than doubles Opus 5's score (58.7% vs. 29.0%), demonstrating a major improvement in agentic scientific research capabilities.
2.4 Important Caveat on Benchmark Interpretation
Anthropic explicitly notes:
"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."
All Opus 5.5 results use adaptive thinking at max effort unless otherwise noted. The model was evaluated with production safeguards enabled, which may reduce scores on tasks where safeguards intervened (cybersecurity tasks fell back to Opus 4.8, biology and frontier LLM development tasks fell back to Opus 5).
3. Pricing & Efficiency: The 40% Cost Reduction
3.1 Price Comparison
| Metric | Opus 5.5 | Opus 5 | Change |
|---|---|---|---|
| Input tokens | $4/M | $5/M | -20% |
| Output tokens | $20/M | $25/M | -20% |
| Cache reads | $0.20/M | $0.50/M | -60% |
| Cache writes | $5/M | $6.25/M | -20% |
The cache read price reduction (60%) is particularly significant for agentic and coding workloads, where cached input tokens typically make up the majority of costs.
3.2 Fast Mode
Opus 5.5 also offers Fast mode in Claude Code and the Claude Platform:
| Metric | Standard | Fast Mode |
|---|---|---|
| Input | $4/M | $8/M |
| Output | $20/M | $40/M |
| Speed | Baseline | Up to 2.5× faster |
Fast mode doubles the price but delivers up to 2.5× speed, making it cost-effective for latency-sensitive interactive workflows.
3.3 Real-World Cost Savings
Anthropic's internal tests show 40% lower costs on typical workloads, driven by both lower per-token prices and fewer tokens per task:
| Task | Opus 5.5 | Opus 5 | Savings |
|---|---|---|---|
| 200K-line codebase audit | <3 hours | >20 hours | 6.7× faster, 2.5× fewer tokens |
| HAProxy C→Rust translation | 9.5 hours | — | 51% less than Fable 5.1 |
| M&A financial model | 63 minutes | 93 minutes | 50% less cost |
| Billing refactor analysis | 63 min, thorough | 93 min, minor errors | Better quality, faster, cheaper |
3.4 Usage Limit Changes
In addition to price reductions, Anthropic is:
- Increasing five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans
- Providing subscription users a rate limit reset that can be saved and used on demand
4. Safety & Alignment: Best-in-Class Scores
4.1 Automated Behavioral Audit
Opus 5.5 achieves the best scores of any model to date on Anthropic's automated behavioral audit — an alignment suite testing Claude across thousands of simulated scenarios:
| Safety Dimension | Opus 5.5 vs. Prior Models |
|---|---|
| Hard-to-reverse actions | Much less likely than recent models |
| Acting outside boundaries | Much less likely than recent models |
| Prompt injection resistance | More resistant than Opus 5 |
| Overall behavioral audit | Best scores of any model tested |
4.2 Prompt Injection Defense
On prompt injection attacks, Opus 5.5 matches or beats Opus 5 in every setting tested — coding, tool use, computer use, and web browsing. On a benchmark run by AI security firm Gray Swan, Opus 5.5 ties Fable 5.1 for the lowest prompt injection success rate of any model tested.
4.3 Enterprise Security Features
For enterprise deployments, Opus 5.5 includes:
- Pre-action classifier — screens every action before it runs
- Open-source sandbox — security teams can audit the execution environment
- Code review — catches vulnerabilities before merge
- Preserved thinking — anti-distillation safeguard preventing API users from editing Claude's prior context to extract reasoning
4.4 Verification Programs
Because Opus 5.5 is comparable to Claude Mythos 5.1 in biology and cybersecurity, it deploys with similar safeguards:
| Program | Access | Status |
|---|---|---|
| Life Sciences Verification Program | Vetted organizations | Available now |
| Cyber Verification Program | Verified cybersecurity practitioners | Expanding in coming weeks |
See Claude Fable 5 1 Mythos 5 1 Scientific Frontier Agentic Coding Enterprise Frontier Safeguards 2026 09 17 for the original Fable/Mythos verification program details.
5. Real-World Testing: What Early Users Found
5.1 Code Migration at Scale
One early tester completed a 680,000-line code migration in less than a day — work that would have taken an engineering team weeks. This demonstrates Opus 5.5's ability to handle large-scale, complex software engineering tasks with minimal human oversight.
5.2 Performance Optimization
When asked to cut load times across every page of a web app:
- Opus 5.5: Succeeded 39 of 40 times
- Opus 5: Made smaller improvements that also altered the app's behavior
5.3 Game Development
In a test where several Claude models built a game from a single prompt, Opus 5.5 scored higher than any other model on the strength of its graphics and polish.
5.4 Knowledge Work: Financial Analysis
Walleye Capital, an investment firm and early tester, reported that Opus 5.5 largely solved their evaluation suite on its lowest setting. On higher settings, it performed even better — notably catching an error in their evaluation instructions that no other model had identified before.
5.5 Code Review
Deloitte Consulting reported:
"Even at its lowest effort setting, Claude Opus 5.5 caught 72% of known bugs in our code reviews to Opus 5's 56% at high effort, with fewer false alarms and a fraction of the output."
5.6 GitHub Integration
Mario Rodriguez, Chief Product Officer at GitHub, noted:
"In our testing across GitHub Copilot CLI and VS Code, Claude Opus 5.5 used among the fewest tokens and steps we measured. In VS Code, it solved more terminal tasks than Opus 5 in less than half the steps."
6. Communication Improvements: "It Writes the Way I Do"
One of the most significant improvements in Opus 5.5 is communication quality — addressing one of the most common complaints about Opus 5:
| Aspect | Opus 5 | Opus 5.5 |
|---|---|---|
| Information prioritization | Buried key findings | Most important info up front |
| Jargon usage | Frequent | Reduced |
| Writing style | Idiosyncratic phrases | Natural, follows user rules |
| Long-session usability | Degrades over time | Better work partner over long sessions |
| Safety through clarity | Harder to verify | Easier to follow and check |
Anthropic demonstrated this with a side-by-side comparison of a billing bug explanation, where Opus 5.5's response was significantly more structured, clearer, and easier to act upon than Opus 5's technically correct but confusing explanation.
"It writes the way I do." — Early tester
This improvement is framed as both a practical benefit (better collaboration) and a safety benefit (easier to verify and check the model's work).
7. The Claude 5.5 Family Roadmap
Opus 5.5 is the first of three planned models in the Claude 5.5 family:
| Model | Status | Expected Timeline |
|---|---|---|
| Claude Opus 5.5 | Released | September 22, 2026 |
| Claude Sonnet 5.5 | Announced | Coming weeks |
| Claude Haiku 5.5 | Announced | Coming weeks |
Anthropic states that Sonnet 5.5 and Haiku 5.5 will include "many of the same improvements to performance, efficiency, and safety" as Opus 5.5. This suggests a comprehensive family-wide upgrade rather than an Opus-specific improvement.
8. Integration with Prior Research
8.1 The September 2026 Anthropic Arc
Opus 5.5 completes a three-part September narrative for Anthropic:
| Date | Event | Significance |
|---|---|---|
| Sep 1 | Fable 5.1 + Mythos 5.1 release | Frontier capability push (Claude Fable 5 1 Mythos 5 1 Scientific Frontier Agentic Coding Enterprise Frontier Safeguards 2026 09 17) |
| Sep 12 | Amodei's pacing essay | Strategic pause call |
| Sep 22 | Opus 5.5 release | Efficiency + safety over raw capability |
This progression — push capability, call for pacing, release efficient model — demonstrates a deliberate strategy of demonstrating that progress doesn't require constant capability escalation.
8.2 Positioning in the Frontier Landscape
| Model | Release | Terminal-Bench 4.0 | GDPval-AA v2.1 | Input Price | Output Price |
|---|---|---|---|---|---|
| Claude Opus 5.5 | Sep 22 | 66.4% | 1846 | $4 | $20 |
| Claude Fable 5.1 | Sep 1 | 55.8% | 1735 | $10 | $50 |
| Claude Opus 5 | Earlier | 52.3% | 1708 | $5 | $25 |
| GPT-6 Astra | Sep | 57.9% | 1542 | $10 | $50 |
| GPT-6 Sol | Sep | — | — | $2 | $10 |
| Grok 4.7 | Sep 21 | — | — | $2 | $6 |
| Step 5 Preview | Sep 20 | 33.3% | 1571 | ~$1 | ~$2.70 |
See Stepfun Step 5 Preview 600b Sparse Moe Agentic Coding Pareto Frontier 2026 09 21 for Step 5 Preview and Qwen3 8 Max First Open Source Flagship Hybrid Deltanet Moe Coding Cowork 2026 09 18 for Qwen3.8-Max.
8.3 The Efficiency Frontier
Opus 5.5 occupies a unique position: it matches or exceeds Fable 5.1 on most benchmarks while costing 40% less. This creates a new efficiency frontier for closed-source models:
| Model | Intelligence (GDPval-AA) | Cost per Task (relative) | Efficiency |
|---|---|---|---|
| Opus 5.5 | 1846 | 1.0× | Baseline |
| Fable 5.1 | 1735 | ~2.5× | Lower |
| Opus 5 | 1708 | ~1.7× | Lower |
| GPT-6 Astra | 1542 | ~5× | Much lower |
| Step 5 Preview | 1571 | ~0.3× | Higher (but lower intelligence) |
9. Key Takeaways
-
Pacing in practice. Opus 5.5 is the concrete demonstration of Amodei's pacing call — a model tested by external evaluators (Frontier Design, METR) with wider safety margins, released not as a capability leap but as an efficiency improvement.
-
Fable 5.1 performance at Opus prices. The most significant achievement is matching Fable 5.1's intelligence at 40% of Opus 5's cost, creating a new sweet spot for enterprise agentic workloads.
-
Safety scores are best-in-class. Opus 5.5 achieves the highest scores on Anthropic's behavioral audit of any model tested, with particular improvements in resisting hard-to-reverse actions and prompt injection.
-
Communication quality is a safety feature. The improvement in writing clarity isn't just about usability — Anthropic explicitly frames it as a safety benefit, making the model's work easier to verify and check.
-
The 5.5 family upgrade is comprehensive. With Sonnet 5.5 and Haiku 5.5 coming in the coming weeks, the efficiency and safety improvements will cascade across Anthropic's entire model lineup.
-
Enterprise readiness is strong. The pre-action classifier, open-source sandbox, code review, and verification programs make Opus 5.5 particularly well-suited for production enterprise deployments.
-
Real-world results exceed benchmark margins. Early testers reported dramatic improvements — 680K-line migrations in a day, 200K-line audits in 3 hours, catching evaluation errors no other model found — suggesting the real-world impact may be larger than benchmark scores indicate.
-
The preserved thinking safeguard extends to Opus. The anti-distillation mechanism previously limited to Fable 5.1 is now available on Opus 5.5, protecting against reasoning extraction attempts.
10. References & Resources
Official Sources
- Anthropic: Introducing Claude Opus 5.5 — Official announcement with benchmarks and pricing
- Claude Opus 5.5 System Card (PDF) — Full technical evaluation and safety analysis
- Anthropic: Claude Fable 5.1 and Mythos 5.1 — Previous generation reference
- Life Sciences Verification Program — Biology research access
- Cyber Verification Program — Cybersecurity access
Related Research in This Journal
- Claude Fable 5 1 Mythos 5 1 Scientific Frontier Agentic Coding Enterprise Frontier Safeguards 2026 09 17 — Fable 5.1 and Mythos 5.1 deep-dive
- Ai News Week 2026 09 14 2026 09 21 — Weekly news covering Amodei's pacing call
- Stepfun Step 5 Preview 600b Sparse Moe Agentic Coding Pareto Frontier 2026 09 21 — Step 5 Preview comparison
- Qwen3 8 Max First Open Source Flagship Hybrid Deltanet Moe Coding Cowork 2026 09 18 — Qwen3.8-Max context
External Benchmarks
- Terminal-Bench 4.0 — Agentic coding benchmark
- GDPval-AA v2.1 — Real-world work across 44 occupations
- OSWorld 2.0 — Computer use benchmark
- Gray Swan — AI security benchmarking
11. Future Directions
What to Watch
-
Sonnet 5.5 and Haiku 5.5 releases: Expected in the coming weeks, these will extend the 5.5 family's efficiency and safety improvements to mid-tier and speed-optimized models.
-
Cyber Verification Program expansion: Currently limited to Opus 5 and Fable 5.1, access to Opus 5.5 for cybersecurity practitioners is coming in the coming weeks.
-
External evaluator findings: The full reports from Frontier Design and METR's pre-release evaluation may provide additional insights into Opus 5.5's safety profile.
-
Anthropic IPO (November target): The Opus 5.5 release timing (one month before the targeted IPO) suggests it may be positioned as a key product for the public offering.
-
Community benchmarks: Independent evaluations of Opus 5.5 beyond Anthropic's own testing will be critical for validating the claimed improvements.
-
Competitive response: The 40% cost reduction may pressure OpenAI and Google to reconsider their pricing strategies, particularly for mid-tier models.
-
Fast mode adoption: The availability of 2.5× speed at 2× price could shift the cost-performance tradeoff for latency-sensitive agentic workloads.
-
Preserved thinking impact: The extension of anti-distillation safeguards to the Opus line may influence how researchers approach model evaluation and distillation.
-
Enterprise adoption: The combination of improved safety, lower cost, and enterprise security features could accelerate Opus 5.5 adoption in regulated industries.
-
Pacing follow-through: Whether Anthropic maintains its pacing commitments in future releases will be a key test of the credibility of the September 12 essay.
Citation:
@misc{claude_opus_5_5,
title = {{Claude Opus 5.5}: The First Model in the Claude 5.5 Family},
url = {https://www.anthropic.com/claude-opus-5-5},
author = {{Anthropic}},
month = {September},
year = {2026}
}
@techreport{claude_opus_5_5_system_card,
title = {{System Card: Claude Opus 5.5}},
url = {https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf},
author = {{Anthropic}},
month = {September},
year = {2026}
}
This article was researched using official Anthropic sources: the Claude Opus 5.5 announcement page, the System Card PDF, the Fable 5.1/Mythos 5.1 announcement, the Life Sciences Verification Program page, and the Cyber Verification Program documentation. All benchmark figures and real-world results are from Anthropic's published evaluations and early tester reports.
🔗 Referenced by
- 🔬Claude Haiku 5.5: Anthropic's Cheapest, Fastest Small Model With Adjustable Effort, 1M Context, and Browser Use2026-10-08T00:00:00.000Z
- 🔬Gemini 4 Argon: Google's Frontier Comeback with 1M Output Tokens, Cyber Defense, and the Fairwind Rollout2026-09-30T00:00:00.000Z
- 🔬Claude Sonnet 5.5: The 60-Point Coding Jump, 30% Speed Gain, and First Sonnet with Cyber Safeguards2026-09-29T00:00:00.000Z
- 🔬OpenAI DevDay 2026: Dots Always-On Agents, GPT-6.1 Sol, Ultrafast, Decisions API, and the Astra Pull2026-09-29T00:00:00.000Z
- 🔬MiniMax M3.1-Flash-Preview: The Quiet Launch, Stealth Benchmarking, and the New Speed Tier for Everyday Coding2026-09-28T00:00:00.000Z
- 🔬Grok 4.7: Long-Running Agents, Frontier Coding at $2/M, and xAI's New Safeguard Stack2026-09-25T00:00:00.000Z
- 🔬GPT-6 Sol and Luna: 50% Price Cuts, Agentic Coding at Scale, and the New Cost-Efficiency Frontier2026-09-24T00:00:00.000Z
- 📚Wiki Index2026-06-17T00:00:00.000Z
- 📚Wiki Log2026-06-17T00:00:00.000Z