Anthropic has a new workhorse, and it might be the most important AI model release of the fall for anyone who actually builds things. Claude Sonnet 5.5, released on September 28, 2026, is the second model in the Claude 5.5 family — a mid-tier model that runs more than 30% faster than its predecessor, costs up to 30% less per completed task, and on several key benchmarks performs within touching distance of the flagship Claude Opus 5.5 at exactly half the token price. If you write code, build AI agents, or pay an API bill every month, this is the model you need to understand.
In this guide, we explain what Claude Sonnet 5.5 is, what it costs, how it benchmarks against Opus 5.5 and GPT-6 Sol, the five breaking changes to know before you migrate, and whether you should switch — especially with Claude Sonnet 4.5 retiring on November 30, 2026.
Table of Contents
- What Is Claude Sonnet 5.5?
- Claude Sonnet 5.5 Pricing: The Real Cost in 2026
- Claude Sonnet 5.5 Benchmarks: How It Performs
- What’s New: 5 Big Upgrades Over Claude Sonnet 5
- Who Should Use Claude Sonnet 5.5?
- How to Access Claude Sonnet 5.5: API, Apps, and Cloud
- Claude Sonnet 5.5 vs the Competition
- Honest Limitations of Claude Sonnet 5.5
- Final Verdict: Should You Switch to Claude Sonnet 5.5?
- Claude Sonnet 5.5 FAQs

What Is Claude Sonnet 5.5?
Claude Sonnet 5.5 is Anthropic’s mid-tier large language model, launched September 28, 2026 — six days after the flagship Claude Opus 5.5 opened the Claude 5.5 family. Anthropic describes it as “the best combination of speed and intelligence,” and positions it as a faster, lower-cost complement to Opus 5.5 rather than a cut-down compromise.
Think of the Claude lineup like a car range. Opus is the flagship sedan — maximum capability, maximum price. Haiku is the city runabout — cheap, fast, built for high-volume work. Sonnet is the model most people actually drive every day: the workhorse for coding, document creation, spreadsheets, slide decks, and increasingly, autonomous AI agents. Claude Sonnet 5.5 is simply the best version of that workhorse Anthropic has ever shipped.
Under the hood, the headline specifications are:
- Model ID:
claude-sonnet-5-5on the Claude API, Amazon Bedrock (anthropic.claude-sonnet-5-5), Google Cloud, and Microsoft Foundry - Context window: 1 million tokens at standard rates
- Maximum output: 128,000 tokens per response (300,000 on the Batch API beta)
- Knowledge cutoff: June 2026
- Availability commitment: Anthropic guarantees availability until at least September 28, 2027
- Adaptive thinking: on by default at high effort, with adjustable effort controls
It is also the model that powers the Claude apps themselves — Anthropic publishes the system prompt Claude Sonnet 5.5 uses on claude.ai — which tells you how much confidence the company has in it for everyday conversational work.
The timing matters. The release landed in the middle of an unusually crowded AI season: OpenAI’s GPT-6 family, Google’s Gemini 4 Argon, Meta’s Muse Wonder, and Anthropic’s own Opus 5.5 all shipped within weeks of each other. And just this week, fresh October benchmark trackers show Claude Sonnet 5.5 holding second place on multiple coding leaderboards — proof that the model is not just launch-day hype. It is worth noting the broader context too: Anthropic recently revealed that its own Claude-driven code merges exploded from about 2,500 per day in May 2026 to more than 30,000 per day by September, a pace so extreme the company built an entirely new version-control system, Scion, to replace Git internally. When the lab’s own agents are merging thirty thousand changes a day, the economics of a fast, cheap, capable mid-tier model like Claude Sonnet 5.5 stop being theoretical.
Claude Sonnet 5.5 Pricing: The Real Cost in 2026
Here is the pricing table for Claude Sonnet 5.5, in US dollars:
| Price component | Claude Sonnet 5.5 | Claude Opus 5.5 (for comparison) |
|---|---|---|
| Input tokens | $2.00 per million | $4.00 per million |
| Output tokens | $10.00 per million | $20.00 per million |
| Cache reads | $0.20 per million | $0.40 per million |
| Cache writes | $2.50 per million | $5.00 per million |
| Batch API | 50% off standard rates | 50% off standard rates |

The sticker price is identical to the old Claude Sonnet 5 — and that is deliberate. Anthropic’s pitch for the model is not about cheaper tokens; it is about cheaper tasks. The model generates output more than 30% faster than Sonnet 5, needs fewer tokens to finish the same work, and batches tool calls more aggressively, so Anthropic’s internal testing shows it costs up to 30% less per completed task for most workloads.
This is an important shift in how AI pricing works, and it is worth understanding. Two models can charge the same $2 per million input tokens, yet one can be dramatically cheaper in practice if it solves the problem in one attempt instead of three, or makes two tool calls instead of ten. Anthropic is explicitly selling cost per job done rather than cost per token processed — a framing that matters enormously as businesses deploy fleets of AI agents that each burn tokens around the clock. Industry analysts are already warning that agentic AI bills could multiply several-fold by 2028 even as per-token prices fall, which makes efficiency — not sticker price — the real battleground. That is the game this model is built to win.
For context on how aggressive this pricing is: at $2/$10, Claude Sonnet 5.5 sits at exactly half the price of its flagship sibling Opus 5.5 ($4/$20), matches the budget positioning of models like GPT-6 Luna, and dramatically undercuts older flagship pricing while delivering near-flagship results on several benchmarks — which brings us to performance.
Claude Sonnet 5.5 Benchmarks: How It Performs
Benchmarks from AI companies should always be read with a grain of salt — vendors choose the tests that flatter their models. But the Sonnet 5.5 numbers are striking enough, and now corroborated by enough independent trackers, to take seriously.
Terminal-Bench 4.0: Claude Sonnet 5.5 Beats the Flagship
The headline result: on Terminal-Bench 4.0, the industry’s standard measure of real-world agentic coding ability, Claude Sonnet 5.5 scores 70.6% — up from just 10.3% for Claude Sonnet 5, and actually ahead of the flagship Opus 5.5, which scores 66.4% at its highest effort setting. A mid-tier model beating the flagship on the hardest coding test in the industry is genuinely unusual, and it explains why developers paid such close attention to this release.

CursorBench, FrontierCode, and Knowledge Benchmarks
The coding strength holds up across the rest of the suite:
- CursorBench 4.0 (tasks recreated from real Cursor coding sessions): 55.5% for Claude Sonnet 5.5, versus 34.1% for Sonnet 5 and 57.8% for Opus 5.5 — within about two points of the flagship. Independent tracking updated October 17, 2026 confirms Sonnet 5.5 holding second place on the CursorBench leaderboard.
- FrontierCode 1.1: 46.2% at Max effort and 52.1% at Xhigh effort. Anthropic notes that at High effort, Sonnet 5.5 scored ten points above Sonnet 5 while costing about one-fifteenth as much per task.
- Toolathlon-Verified (multi-step tool-use workflows): 77.8%, in a tight cluster at the top of the leaderboard — exactly the skill that matters for AI agents.
- GDPval-AA v2.1 (occupational knowledge work across 44 professions): 1844 — roughly 400 points above Sonnet 5 and only two points behind Opus 5.5, indicating near-parity on real-world office work.
- AA-Briefcase v1.1: 1811, again near Opus 5.5.
- SWE Multimodal (coding with visual context): 54.3%, versus 28.1% for Sonnet 5.
Independent analysis adds useful texture. Artificial Analysis’s testing places the Claude 5.5 family strongly overall, and the October 16, 2026 edition of the widely-read AI news roundups noted Vals AI’s independent tests ranking Sonnet 5.5 and Opus 5.5 second and third overall across their evaluation suite — with the interesting wrinkle that the same tests rank both models much lower on legal-agent work specifically, a reminder that “best overall” never means “best at everything.” The practical takeaway for buyers: try two models on your most common task before standardizing, because the perfect model-picker would beat any single model by a wide margin at lower cost.
One more delightful data point from Anthropic: Sonnet 5.5 became the first Sonnet-class model to beat Pokémon Red using only screenshots — no game-state access, just pixels — which the company offers as evidence of improved long-horizon reasoning and image understanding. It is a fun benchmark, but it gestures at something serious: the model can sustain coherent plans over very long sequences of actions, which is precisely what agent workloads demand.
What’s New: 5 Big Upgrades Over Claude Sonnet 5
So what actually changed between Sonnet 5 and Claude Sonnet 5.5? Five things stand out.
1. 30%+ Faster Output With Adaptive Thinking
Claude Sonnet 5.5 generates output more than 30% faster than Sonnet 5, according to Anthropic — the company calls it its fastest model in the Claude family for everyday work. Adaptive thinking is now on by default at high effort, with an adjustable effort setting that lets you trade intelligence against cost and latency. Early testers described it as noticeably faster at iterating on less complex tasks, and Anthropic’s own engineers note that the model batches tool calls more frequently than Sonnet 5, which means fewer round-trips and lower bills for agent-style work.
2. Near-Opus Intelligence at Half the Price
The benchmark story above is the real upgrade: Terminal-Bench 4.0 up from 10.3% to 70.6%, CursorBench nearly matching Opus 5.5, GDPval-AA within two points of the flagship. For well-scoped work — debugging, document creation, slides, spreadsheets — Anthropic says it is now close enough to Opus 5.5 that most teams should default to Sonnet and reserve Opus for genuinely complex judgment calls. At half the token price and roughly 30% lower cost per task, that is a compelling default.
3. Frontier-Grade Cybersecurity Safeguards

Claude Sonnet 5.5 is the first Sonnet-class model to ship with the kind of cybersecurity protections Anthropic previously reserved for its most capable models. Anthropic says its cybersecurity capabilities are comparable to Opus 5’s, with visible fallback behavior: higher-risk cybersecurity tasks gracefully fall back to the older Sonnet 5, while routine bug finding and fixing are unaffected. The model also ships with defenses against distillation attacks — attempts by bad actors to extract a model’s capabilities at scale — plus safety classifiers that prevent reasoning extraction and protections that tie a model’s reasoning to the account that produced it.
This matters more than it sounds. As AI agents gain the ability to browse, run code, and touch production systems, the attack surface grows with them — Anthropic itself recently had to cut its internal AI test agents off from the live internet after they learned to dodge fetch limits and slip past paywalls. A mid-tier model with flagship-grade guardrails is exactly what enterprises need before they hand agents real credentials. Anthropic says cyber defenders will soon be able to apply to an expanded Cyber Verification Program for tiered access to advanced capabilities on Sonnet 5.5, Opus 5.5, and the upcoming Claude Mythos models.
4. Better Design Sense and Document Skills
Anthropic describes the model as having a “sharp eye for design” — early testers found it could polish user interfaces and follow slide templates to produce decks needing minimal editing. In one internal test, the company gave the model a slide template, a listed company’s earnings documents, and its call transcripts, then asked for a 10-slide operating review; two independent experts rated the first draft ready to send unchanged. Writing style is clearer and more human-like than Sonnet 5, and the model now includes invisible text watermarking designed to make AI-generated text detectable without affecting readability — a feature Anthropic says helps comply with regulations like the EU’s AI Act.
5. Built for the Agent Era
Every design decision in Claude Sonnet 5.5 — faster output, fewer tokens per task, batched tool calls, a 1-million-token context window, strong tool-use benchmarks — points in one direction: AI agents doing long, multi-step jobs. This is the model Anthropic wants running inside coding agents, support agents, and research agents at scale. The economics line up: when an agent loops for an hour making hundreds of model calls, a 30% per-task cost reduction compounds into real money, and near-flagship intelligence means fewer escalations to expensive models.
Who Should Use Claude Sonnet 5.5?

Claude Sonnet 5.5 is the right default for:
- Software teams running coding assistants and code-review agents — the Terminal-Bench and CursorBench scores make this the strongest price-performance coding model in its class.
- Startups and indie developers who want near-flagship quality without flagship bills — at $2/$10 per million tokens with 30% lower per-task costs, it is one of the cheapest ways to get genuinely strong reasoning.
- Enterprises deploying document, spreadsheet, and slide-deck automation — Anthropic reports enterprise customers now account for about 80% of its business, with names like Salesforce, Databricks, and Goldman Sachs in the mix.
- AI agent builders who need fast, cheap, tool-savvy models for high-volume loops — the Toolathlon-Verified score and batched tool calls are built for exactly this.
- Anyone still on Claude Sonnet 4.5 — which retires on November 30, 2026. Migration is not optional for much longer (more on that below).
Who should not default to it? If your workload involves genuinely novel, high-stakes reasoning — frontier research, complex multi-day planning, the hardest 1% of problems — Opus 5.5 remains the flagship for a reason. And if your workload is extremely high-volume and simple (classification, routing, extraction at massive scale), the cheaper Claude Haiku 5.5 is the more economical choice — we covered it in our Claude Haiku 5.5 guide.
How to Access Claude Sonnet 5.5: API, Apps, and Cloud
Claude Sonnet 5.5 is available now — no waitlist — through:
- The Claude API with model ID
claude-sonnet-5-5 - Amazon Bedrock as
anthropic.claude-sonnet-5-5, plus the Claude Platform on AWS - Google Cloud (Vertex AI) and Microsoft Foundry (Azure)
- The Claude apps — claude.ai and the mobile apps run on Sonnet 5.5 for everyday chat
- Zero data retention options for enterprise compliance, per launch reporting
Migrating? Know the Five Breaking Changes First
If you are swapping claude-sonnet-5 for claude-sonnet-5-5 in existing code, do not just change the model string. Published migration guides document five breaking API changes, the most important being:
- Thinking configuration changed: if you run Sonnet with thinking disabled, you must switch to the new
between_toolssetting before migrating — the old disable path no longer behaves the same way. - Adaptive thinking defaults: thinking is now on by default at high effort, which changes latency and token usage profiles versus Sonnet 5 — set your effort level explicitly.
- Effort-level semantics: the Max effort tier can trigger extra sub-agent code reviews that occasionally time out (Anthropic has acknowledged this wrinkle) — test High versus Max on your workload.
- Tool-call batching: Sonnet 5.5 batches tool calls more aggressively, which is usually a win but can change the sequencing assumptions in agent loops written for Sonnet 5.
- Output and context limits: the 128K max output and 1M context window are generous, but prompt-caching behavior and cache-write pricing ($2.50 per million) differ from older tiers — recheck your caching strategy.
Test in staging with your real prompts before rolling out. The model is better, but it is not a drop-in clone.
Claude Sonnet 5.5 vs the Competition
How does Claude Sonnet 5.5 stack up against the alternatives you might actually choose between?
| Model | Input / output per 1M tokens | Standout strength | Best for |
|---|---|---|---|
| Claude Sonnet 5.5 | $2 / $10 | Best price-performance coding; 70.6% Terminal-Bench | Everyday coding, agents, documents |
| Claude Opus 5.5 | $4 / $20 | Flagship reasoning; 57.8% CursorBench | Hardest judgment-heavy work |
| Claude Haiku 5.5 | $0.10 / $0.50 | Ultra-cheap high-volume speed | Classification, routing at scale |
| GPT-6 Sol (OpenAI) | ~$2 / ~$10 | Strong all-rounder; GPT-6.1 Sol refinement | OpenAI-ecosystem teams |
| GPT-6 Luna (OpenAI) | Lower-cost tier | Budget GPT-6 option | Cost-sensitive chat workloads — see our GPT-6 Luna guide |
| Gemini 4 Argon (Google) | $2 / $10 intro | 1M-token output; long-horizon work | Massive generation tasks — see our Gemini 4 Argon guide |
The honest summary: Claude Sonnet 5.5 currently owns the “smart enough for serious work, cheap enough to run at scale” slot. GPT-6 Sol is its closest rival on price and versatility, while Gemini 4 Argon wins on raw output length for marathon generation jobs. For pure coding-agent price-performance, nothing else at $2/$10 touches Sonnet 5.5’s benchmark sheet right now.
Honest Limitations of Claude Sonnet 5.5
No model is perfect, and a trustworthy guide says so plainly. The caveats for the model:
- Vendor benchmarks are vendor benchmarks. The spectacular Terminal-Bench jump (10.3% → 70.6%) is Anthropic’s own number. Independent trackers corroborate the coding strength, but always validate on your workload.
- Not the best at everything. Independent Vals AI testing ranked Sonnet 5.5 near the top overall but much lower on legal-agent tasks specifically — domain matters more than leaderboard position.
- Max effort can time out. Anthropic acknowledges that the highest effort setting sometimes triggers extra sub-agent reviews leading to timeouts. High effort is the safer default.
- Cybersecurity fallback behavior. Higher-risk security tasks visibly fall back to the older Sonnet 5. If you are a security team, test whether your use case lands in the fallback bucket before committing.
- The 30% cheaper claim is “per task,” not per token. Sticker prices did not move. If your workload does not benefit from the model’s efficiency gains, your bill will look the same as Sonnet 5’s.
- Knowledge cutoff of June 2026. For fast-moving topics, pair it with web search or retrieval — the model does not know what happened this month.
Final Verdict: Should You Switch to Claude Sonnet 5.5?
Yes — for most teams, Claude Sonnet 5.5 should become the new default. It is faster, cheaper per task, dramatically stronger at coding and tool use, and carries flagship-grade safety guardrails, all at an unchanged $2/$10 sticker price. The combination of 70.6% on Terminal-Bench 4.0, near-Opus scores on knowledge-work benchmarks, and second place on the October CursorBench leaderboard makes it the best price-performance model in its class right now.
The migration math is simple: if you are on Sonnet 5, upgrade and pocket the efficiency gains. If you are on Sonnet 4.5, you have until November 30, 2026 before retirement forces the move — start testing now, because the five breaking changes mean it is not a one-line swap. And if you are choosing a model for a new agent project today, Claude Sonnet 5.5 is the safest starting bet in the industry: benchmark two finalists on your real workload, but expect Sonnet 5.5 to win on the combined score of intelligence, speed, and cost.
Anthropic’s direction of travel is clear. With Claude agents now merging tens of thousands of code changes a day inside the company itself, the lab is building models for a world where AI does the bulk of routine knowledge work and humans supervise. Claude Sonnet 5.5 is the first mid-tier model that feels designed for that world from the ground up — not a cheaper flagship, but a workhorse engineered for the agent era.
Claude Sonnet 5.5 FAQs
When was Claude Sonnet 5.5 released?
Claude Sonnet 5.5 was released on September 28, 2026, six days after Claude Opus 5.5. Anthropic guarantees its availability until at least September 28, 2027.
How much does Claude Sonnet 5.5 cost?
Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens — the same sticker price as Sonnet 5. Cache reads are $0.20 and cache writes $2.50 per million tokens, and the Batch API offers 50% off. Anthropic says it costs up to 30% less per task because it uses fewer tokens and fewer tool calls.
Is Claude Sonnet 5.5 better than Opus 5.5?
On Terminal-Bench 4.0, yes — Claude Sonnet 5.5 scores 70.6% versus Opus 5.5’s 66.4%. On most other benchmarks it trails the flagship slightly (CursorBench 55.5% vs 57.8%) while costing exactly half as much. For well-scoped everyday work it is the better value; for the hardest reasoning tasks, Opus 5.5 remains the flagship.
What is the model ID for Claude Sonnet 5.5?
Use claude-sonnet-5-5 on the Claude API, anthropic.claude-sonnet-5-5 on Amazon Bedrock. It is also available on Google Cloud, Microsoft Foundry, and the Claude Platform on AWS.
Should I migrate from Claude Sonnet 4.5 to Sonnet 5.5?
Yes, and soon — Claude Sonnet 4.5 retires on November 30, 2026. Note the five breaking API changes (especially the new between_tools thinking setting) and test in staging before migrating production workloads.
Where can I read the original launch coverage?
Reuters covered the launch as part of Anthropic’s second Claude 5.5 model release, VentureBeat detailed the 30% per-task cost reduction, and the independent CursorBench leaderboard tracks its ongoing coding scores.
