Claude Sonnet 5.5: The 60-Point Coding Jump, 30% Speed Gain, and First Sonnet with Cyber Safeguards
On September 28, 2026, Anthropic released Claude Sonnet 5.5 — the second model in the 5.5 family. At the same price as Sonnet 5 ($2/$10 per million tokens), it scores 70.6% on Terminal-Bench 4.0 (vs. Sonnet 5's 10.3%), runs 30%+ faster, costs up to 30% less per task, and is the first Sonnet model to launch with cyber safeguards and distillation protection.
Claude Sonnet 5.5: The 60-Point Coding Jump, 30% Speed Gain, and First Sonnet with Cyber Safeguards
Executive Summary
On September 28, 2026, Anthropic released Claude Sonnet 5.5, the second model in its new Claude 5.5 family — just six days after launching Claude Opus 5.5. This release completes the initial wave of Anthropic's 5.5 generation, with Claude Haiku 5.5 expected "in the coming weeks."
Sonnet 5.5 is positioned as the workhorse of the family: where Opus 5.5 handles complex work requiring sustained judgment, Sonnet 5.5 dominates well-scoped everyday tasks, bug fixing, document creation, and design work. The headline metric is a 60-point jump on Terminal-Bench 4.0 — from Sonnet 5's 10.3% to 70.6% — making it the most dramatic single-release improvement in agentic coding performance seen from any Sonnet model.
Critically, this performance leap comes at no additional cost. Sonnet 5.5 maintains Sonnet 5's pricing of $2 per million input tokens and $10 per million output tokens, but typically completes tasks using far fewer tokens, resulting in up to 30% lower cost per task. It also generates outputs 30%+ faster than its predecessor, making it Anthropic's fastest Sonnet model to date.
The release introduces significant safety infrastructure: Sonnet 5.5 is the first Sonnet model to launch with cyber safeguards (falling back to Sonnet 5 on high-risk cybersecurity tasks), distillation protection via safety classifiers that prevent reasoning extraction, and expanded preserved thinking that ties Claude's reasoning to the account that created it.
1. The 5.5 Family: Completing the Wave
1.1 Release Timeline
Anthropic's 5.5 family rollout compressed three major releases into a single week:
| Date | Model | Positioning |
|---|---|---|
| Sep 22 | Claude Opus 5.5 | Complex work, careful judgment |
| Sep 28 | Claude Sonnet 5.5 | Everyday tasks, coding, design |
| Coming weeks | Claude Haiku 5.5 | High-volume, cost-sensitive |
This contrasts with previous generation rollouts, where Sonnet and Haiku updates typically arrived weeks or months after Opus. The accelerated pace signals Anthropic's confidence in the 5.5 architecture and a desire to establish the full family before competitors respond.
1.2 Strategic Positioning
Anthropic explicitly frames Sonnet 5.5 as a complement to Opus 5.5, not a replacement:
"Where Opus 5.5 is built for complex work requiring careful judgment, Sonnet 5.5 is strongest at well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets."
This is a deliberate tiering strategy. At lower effort settings (Low, Medium), Sonnet 5.5 costs less per task than Opus 5.5 while delivering competitive results on many benchmarks. At higher settings (High, Max), Sonnet 5.5 can perform comparably to Opus 5.5 on specific tasks at similar cost — but Opus 5.5 remains "clearly stronger at complex, open-ended work requiring sustained judgment."
2. Performance: The 60-Point Coding Leap
2.1 Agentic Coding Benchmarks
The most dramatic improvement is in agentic coding, where Sonnet 5.5 achieves near-Opus-level performance:
| Benchmark | Sonnet 5.5 | Sonnet 5 | Opus 5.5 | GPT-6 Sol |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 70.6% | 10.3% | 66.4% | — |
| FrontierCode 1.1 (Max) | 46.2% | 42.4% | 54.4% | 49.3% |
| FrontierCode 1.1 (XHigh) | 52.1% | — | — | — |
| CursorBench 4.0 | 55.5% | 34.1% | 57.8% | — |
Key observations:
- Terminal-Bench 4.0: Sonnet 5.5 doesn't just improve — it transforms. The 60-point jump (10.3% → 70.6%) puts it ahead of Opus 5.5 on this specific benchmark, suggesting the model has fundamentally better tool-use and multi-step execution capabilities.
- CursorBench 4.0: At 55.5%, Sonnet 5.5 is within 2.3 points of Opus 5.5 (57.8%), making it the closest Sonnet-to-Opus gap on any coding benchmark.
- FrontierCode 1.1: At High effort, Sonnet 5.5 scores 10 points higher than Sonnet 5 at the same setting, at approximately one-fifteenth of the cost per task.
2.2 Knowledge Work
Sonnet 5.5 shows massive gains in real-world knowledge work, nearly matching Opus 5.5:
| Benchmark | Sonnet 5.5 | Sonnet 5 | Opus 5.5 | GPT-6 Sol |
|---|---|---|---|---|
| GDPval-AA v2.1 | 1,844 | 1,449 | 1,846 | 1,487 |
| AA-Briefcase v1.1 | 1,811 | 1,359 | 1,822 | 1,483 |
| Humanity's Last Exam | 64.5% (w/ tools) | 54.9% (w/ tools) | 67.7% (w/ tools) | — |
On GDPval-AA, which tests models across 44 occupations and 9 major industries, Sonnet 5.5 scores nearly level with Opus 5.5 (1,844 vs. 1,846) and ~400 points above Sonnet 5. On AA-Briefcase, which tests long-horizon office work, the gap is similarly narrow.
2.3 Computer Use & Visual Understanding
| Benchmark | Sonnet 5.5 | Sonnet 5 | Opus 5.5 | GPT-6 Sol |
|---|---|---|---|---|
| OSWorld 2.1 (partial) | 80.1% | 57.0% | 81.8% | — |
| Chartography (no tools) | 61.6% | 15.6% | 64.4% | 53.6% |
Two notable achievements:
- Pokémon Red: Sonnet 5.5 is the first Sonnet model to beat Pokémon Red working only from screenshots — a significant milestone in visual game-play capability.
- Chartography: The 46-point improvement over Sonnet 5 (15.6% → 61.6%) represents a fundamental improvement in visual chart recognition, not just incremental gains.
3. Cost and Speed: More Performance, Less Cost
3.1 Pricing
Sonnet 5.5 maintains Sonnet 5's pricing while delivering more capability:
| Price per 1M tokens | Sonnet 5.5 | Sonnet 5 | Opus 5.5 |
|---|---|---|---|
| Cache reads | $0.20 | $0.20 | $0.20 |
| Cache writes | $2.50 | $2.50 | $5.00 |
| Input tokens | $2.00 | $2.00 | $4.00 |
| Output tokens | $10.00 | $10.00 | $20.00 |
3.2 Efficiency Gains
The cost advantage comes from token efficiency, not price reduction:
- Up to 30% less per task: Sonnet 5.5 typically needs far fewer tokens to complete the same work as Sonnet 5.
- 30%+ faster output generation: Measured wall-clock speed improvement over Sonnet 5.
- Fewer tool calls: Early testers observed that Sonnet 5.5 "batched tool calls together more than Sonnet 5, leading to fewer steps and lower costs."
3.3 Effort-Level Economics
The effort ladder enables fine-grained cost control:
| Effort Level | Best For | Relative Cost |
|---|---|---|
| Low | Routine work, quick answers | Lowest |
| Medium (default in apps) | Standard tasks | Low-moderate |
| High (default on Platform) | Complex reasoning | Moderate-high |
| Max | Difficult problems | Highest |
At Low or Medium effort, Sonnet 5.5 often beats Sonnet 5's best score for approximately one-tenth of the cost per task. At higher settings, it can perform comparably to Opus 5.5 at similar cost on specific benchmarks.
4. Real-World Validation: Epic Games and Slack
4.1 Epic Games
Daniel Vogel, COO at Epic Games, reported:
"In Epic's early testing, Claude Sonnet 5.5 cleared the same quality bar you'd expect from a higher-tier model, holding up on a system design audit and a data flow review. The new model managed tens of thousands of lines of code for gameplay system architecture, kept responses snappy, handled multi-hour tasks, and delivered with less prescriptive prompting."
Key takeaways:
- Quality bar: Cleared the threshold typically reserved for higher-tier (Opus-class) models.
- Scale: Handled tens of thousands of lines of code for gameplay system architecture.
- Duration: Managed multi-hour tasks without degradation.
- Prompting: Required less prescriptive prompting than Sonnet 5.
4.2 Slack
Curtis Allen, Principal Engineer at Slack, reported:
"Without changing any of our prompts, Claude Sonnet 5.5 did better than Sonnet 5 on almost all of our offline Slackbot evals, in fewer steps and with about 14% fewer output tokens. When someone gives Slackbot a task, quality and speed are what matter most, and Sonnet 5.5 allows Slackbot to deliver better outcomes for users, faster."
Key takeaways:
- Zero prompt changes: Drop-in improvement without any prompt engineering.
- 14% fewer output tokens: Direct cost savings on production workloads.
- Fewer steps: More efficient tool-use and task completion.
5. Safety: First Sonnet with Cyber Safeguards
5.1 Cybersecurity Safeguards
Sonnet 5.5's cybersecurity capabilities represent a large improvement over Sonnet 5, so Anthropic is deploying it with safeguards similar to those on Opus 5.5:
- Visible fallback: Higher-risk cybersecurity tasks will visibly fall back to Sonnet 5.
- Routine development unaffected: Users can still find and fix bugs in their code as part of normal software development.
- Cyber Verification Program: Cyberdefenders can apply for tiered access to more advanced capabilities on Sonnet 5.5, Opus 5.5, and Claude Mythos models.
This is a significant policy shift: for the first time, a Sonnet model carries the same class of cyber safeguards as an Opus model, acknowledging that the capability gap between tiers has narrowed.
5.2 Distillation Protection
Sonnet 5.5 introduces two distillation countermeasures:
- Safety classifiers: The model includes classifiers that prevent reasoning extraction at industrial scale, blocking the pattern where attackers use thousands of fake accounts to extract a model's capabilities.
- Expanded preserved thinking: Claude's thinking blocks are now tied to the account that created them and cannot be decoupled. Moving conversations between accounts (including switching accounts mid-session in Claude Code) requires following specific migration procedures.
"Most developers won't notice a change. If you move conversations between accounts, including switching accounts mid-session in Claude Code, our docs article explains the change."
5.3 Biology Safeguards
Biology safeguards remain the same as Sonnet 5: they target harmful requests while leaving most research, education, and clinical work unaffected. Organizations can apply to the Life Sciences Verification Program for broader access.
5.4 Alignment Assessment
On Anthropic's automated behavioral audit (~1,850 scenarios):
- Sonnet 5.5 improves on or matches Sonnet 5 on most measures of alignment, resistance to misuse, and honesty.
- On containment evaluations, Sonnet 5.5 comes close to Opus 5.5 in how rarely it tries to escape its sandbox.
- It is the least likely of any Anthropic model to probe the limits of its containers.
- No evidence that Sonnet 5.5 pursues goals conflicting with the user's intention.
6. Migration Guide: Breaking Changes
6.1 What Changes
Sonnet 5.5 introduces several breaking API changes compared to Sonnet 5 and earlier models:
| Setting | Sonnet 5 / Earlier | Sonnet 5.5 |
|---|---|---|
| Thinking budgets | Supported | Returns 400 error |
| Sampling parameters | Supported | Returns 400 error |
| Assistant prefill | Supported | Returns 400 error |
| Forced tool choice | Supported | Returns 400 error |
thinking: {"type": "disabled"} | Supported | Returns 400 error |
| Thinking default | Off without thinking field | Adaptive thinking on by default |
6.2 Turning Off Thinking
To run without up-front thinking on Sonnet 5.5, use the new between_tools setting instead of thinking: {"type": "disabled"}:
# Old (Sonnet 5) — no longer works on Sonnet 5.5
client.messages.create(
model="claude-sonnet-5-5",
thinking={"type": "disabled"}, # Returns 400 error
...
)
# New (Sonnet 5.5) — use between_tools
client.messages.create(
model="claude-sonnet-5-5",
thinking={"type": "between_tools"}, # Keeps up-front thinking off
...
)
6.3 Automated Migration
Anthropic provides a built-in migration tool in Claude Code:
/claude-api migrate this project to claude-sonnet-5-5
This invokes the bundled Claude API skill, which applies the model ID swap, breaking parameter changes, prefill replacement, and effort calibration across your codebase. It also detects Amazon Bedrock and Claude Platform on AWS clients and adjusts model ID formats accordingly.
6.4 Basic Request Example
client = anthropic.Anthropic()
response = client.messages.create(
model="claude-sonnet-5-5",
max_tokens=4096,
messages=[
{"role": "user", "content": "Analyze the trade-offs between microservices and monolithic architectures"}
],
output_config={"effort": "medium"},
)
for block in response.content:
if block.type == "text":
print(block.text)
7. Availability and Deployment
7.1 Platforms
Sonnet 5.5 is available across all major deployment paths:
| Platform | Availability | Model ID |
|---|---|---|
| Claude Platform | Immediate | claude-sonnet-5-5 |
| Amazon Web Services | Immediate | Via Bedrock |
| Google Cloud | Immediate | Via Vertex AI |
| Microsoft Azure | Immediate | Via Azure AI |
| Claude Apps | Immediate | Default in apps |
| Claude Code | Immediate | Default model |
7.2 Zero Data Retention
Like Opus 5.5 and Sonnet 5, Sonnet 5.5 is available with zero data retention, meaning Anthropic does not store input or output data for model improvement.
7.3 Context Window and Output
| Property | Value |
|---|---|
| Context window | Up to 1 million tokens |
| Maximum output | 128,000 tokens (via extended thinking) |
| Input modalities | Text, Image |
| Output modalities | Text |
| Reasoning | Adaptive thinking on by default |
| Effort levels | Low, Medium, High, Max |
8. Connection to Prior Work
8.1 The September 2026 Frontier Landscape
Sonnet 5.5 completes the September 2026 frontier model survey:
- Gpt 6 Sol Luna 50 Percent Price Cut Agentic Coding Cost Efficiency Frontier 2026 09 24 — OpenAI's GPT-6 Sol/Luna with 50% price cuts
- Claude Opus 5 5 Fable 5 1 Performance 40 Percent Cost Reduction Pacing Frontier Safeguards 2026 09 23 — Anthropic's Opus 5.5 with Fable 5.1 performance
- Grok 4 7 Long Running Agents Coding Knowledge Work New Safeguard Stack 2026 09 25 — xAI's Grok 4.7 with multi-hour task training
- Minimax M3 1 Flash Preview Quiet Launch Coding Agent Stealth Benchmark 2026 09 28 — MiniMax's stealth coding model launch
- This article — Anthropic's Sonnet 5.5 with the 60-point coding jump
8.2 The Price-Performance Convergence
The September releases reveal a structural shift: the gap between "premium" and "standard" tier models is collapsing. Sonnet 5.5 at 55.5% on CursorBench is within 2.3 points of Opus 5.5 (57.8%), while costing half as much per token. Combined with GPT-6 Luna at $0.10/$0.50 per million tokens and Grok 4.7 at $2/$6, the price-performance frontier is becoming increasingly crowded.
This convergence forces a strategic question for developers: when a "standard" model performs within 5% of a "premium" model at half the cost, the premium tier must justify its existence through truly unique capabilities — not just incremental improvements.
9. Key Takeaways
-
The coding jump is real: A 60-point improvement on Terminal-Bench 4.0 (10.3% → 70.6%) is not incremental — it represents a fundamental capability shift in Sonnet's agentic coding abilities.
-
More performance, less cost: At the same per-token price as Sonnet 5, Sonnet 5.5 costs up to 30% less per task due to token efficiency. This is the most significant cost-per-task improvement in the Sonnet line.
-
Safety infrastructure matures: The first Sonnet model with cyber safeguards and distillation protection signals that Anthropic treats the Sonnet tier as a potential attack surface, not just a cost-efficient option.
-
Breaking changes require migration: Five API settings now return 400 errors. The
between_toolsthinking setting replacesthinking: {"type": "disabled"}. Use the built-in migration tool in Claude Code. -
The 5.5 family is nearly complete: With Haiku 5.5 expected in the coming weeks, Anthropic will have a full three-tier 5.5 family, covering every price-performance point from high-volume cost-sensitive work to complex judgment tasks.
10. Future Directions
10.1 Haiku 5.5
The most immediate question is what Haiku 5.5 will bring. Given the pattern established by Sonnet 5.5 — dramatic performance gains at unchanged pricing — Haiku 5.5 could establish a new price floor for frontier-adjacent models. If it matches or beats GPT-6 Luna's $0.10/$0.50 pricing while delivering 5.5-generation quality, it would reshape the economics of high-volume AI deployment.
10.2 The Distillation Arms Race
Sonnet 5.5's distillation protection — safety classifiers and account-bound thinking — represents the first significant countermeasure against industrial-scale reasoning extraction. Expect competitors to respond with their own protections, and attackers to adapt. This is the beginning of a new security domain specific to reasoning models.
10.3 The Sonnet-to-Opus Convergence
If Sonnet 5.5 at Max effort can match Opus 5.5 on specific benchmarks at similar cost, the long-term trajectory suggests the premium tier may become increasingly specialized — reserved for tasks that genuinely require sustained judgment across ambiguous, open-ended problems, rather than serving as a general-purpose upgrade.
10.4 Enterprise Adoption
The Epic Games and Slack testimonials suggest enterprise adoption will be rapid. The combination of drop-in improvement (no prompt changes needed), cost reduction (14% fewer tokens at Slack), and multi-hour task capability makes Sonnet 5.5 a compelling upgrade for any organization running Sonnet 5 in production.
11. References & Resources
Official Sources
- Claude Sonnet 5.5 Announcement — Official product page with benchmarks and pricing
- Claude Sonnet 5.5 System Card — Technical evaluation details (PDF)
- Migration Guide — Breaking changes and code migration
- Preserved Thinking Documentation — Account-bound thinking blocks
- Cyber Verification Program — Expanded cyber access for defenders
- Life Sciences Verification Program — Biology safeguards for research organizations
Related Articles
- Claude Opus 5 5 Fable 5 1 Performance 40 Percent Cost Reduction Pacing Frontier Safeguards 2026 09 23
- Gpt 6 Sol Luna 50 Percent Price Cut Agentic Coding Cost Efficiency Frontier 2026 09 24
- Grok 4 7 Long Running Agents Coding Knowledge Work New Safeguard Stack 2026 09 25
- Minimax M3 1 Flash Preview Quiet Launch Coding Agent Stealth Benchmark 2026 09 28
- Ai News Week 2026 09 21 2026 09 28
This article was research-based and sourced exclusively from official Anthropic announcements, system cards, and documentation. All benchmark figures are cited directly from Anthropic's published materials.
🔗 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
- 🔬Grok 4.7: xAI's Frontier Coding and Knowledge Work Model with New Safeguard Stack and Longer RL Training2026-10-06T00:00:00.000Z
- 🔬Qwen3.8-Omni-Flash: Alibaba's Native Omnimodal Model That Turns Media Into Agentic Workflows2026-10-05T00:00:00.000Z
- 🔬Pareto 26.10 Preview: Unbiased's Blended Model That Sets the Cost-Performance Frontier2026-10-02T00:00:00.000Z
- 🔬Gemini 4 Argon: Google's Frontier Comeback with 1M Output Tokens, Cyber Defense, and the Fairwind Rollout2026-09-30T00:00:00.000Z
- 📅September 29: Sonnet 5.5's 60-Point Coding Leap and the 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