Claude Sonnet 5.5 Is Here: Faster, Cheaper, and Nearly Opus-Level
Jaweed Shuja•Sep 29th 26
Every few months a new AI model shows up with "the best ever" on the label, and most of the time the real-world change is small. Claude Sonnet 5.5, which Anthropic released on September 28, 2026, looks like one of the exceptions.
The headline: it's faster than Sonnet 5, it costs less per task, and on several benchmarks it scores the same as or better than Claude Opus 5.5, Anthropic's flagship model, at half the token price.
If you build with the Claude API, run coding agents, or just want to know which model to pick, here's what actually changed.
TL;DR: Claude Sonnet 5.5 at a Glance
- Model ID:
claude-sonnet-5-5 - Speed: generates output 30%+ faster than Sonnet 5 (Anthropic calls it its fastest Sonnet yet)
- Cost: same token prices as before, but up to 30% cheaper per task because it gets there in fewer steps
- Standout result: 70.6% on Terminal-Bench 4.0, ahead of Opus 5.5's 66.4%
- Pricing: $2 per million input tokens, $10 per million output tokens
- Available on: Claude apps, Claude Platform, AWS, Google Cloud, and Microsoft Azure
Why This Release Matters
For a long time the choice between models was simple: pay more for the smartest one, or pay less and accept worse results.
Sonnet 5.5 makes that choice harder. Anthropic describes it as "a faster, lower-cost complement to Claude Opus 5.5," and the numbers back that up. For many everyday workloads, including coding agents, support automation, and computer use, the gap between Sonnet and Opus is now small enough that the price difference probably matters more.
The per-task savings are worth a closer look. The token price didn't drop. What changed is that the model uses fewer tokens and fewer tool calls to finish the same job. In production, that's the number that ends up on your bill.
Claude Sonnet 5.5 Benchmarks: How It Stacks Up
Here are the numbers Anthropic published, compared with Sonnet 5 and Opus 5.5:
| Benchmark | What it tests | Sonnet 5.5 | Sonnet 5 | Opus 5.5 |
|---|---|---|---|---|
| Terminal-Bench 4.0 | Agentic coding | 70.6% | 10.3% | 66.4% |
| FrontierCode 1.1 | Hard coding problems | 46.2% | 42.4% | 54.4% |
| CursorBench 4.0 | Coding in the IDE | 55.5% | 34.1% | 57.8% |
| OSWorld 2.1 | Computer use | 80.1% | 57.0% | 81.8% |
| Humanity's Last Exam | Expert-level reasoning | 64.5% | 54.9% | 67.7% |
| GDPval-AA v2.1 | Real-world knowledge work (Elo) | 1844 | 1449 | 1846 |
| AA-Briefcase v1.1 | Business tasks (Elo) | 1811 | 1359 | 1822 |
| Chartography | Reading charts (no tools) | 61.6% | 15.6% | 64.4% |
What the numbers tell us
1. It beats Opus 5.5 on agentic coding. The Terminal-Bench 4.0 score of 70.6% is the stat that will get the most attention, since it puts a mid-tier model ahead of the flagship on multi-step work in the terminal.
2. It's within a point or two of Opus almost everywhere else. On GDPval-AA the gap is just 2 Elo points (1844 vs. 1846). On OSWorld it's 1.7 percentage points. For most teams that difference won't show up in practice.
3. Opus still leads on the hardest problems. FrontierCode is where the gap is widest: 54.4% for Opus against 46.2% for Sonnet 5.5. If your work involves genuinely difficult, novel engineering problems, Opus is still the safer choice.
4. Vision took a big step forward. Chart reading went from 15.6% to 61.6%. If you've ever watched a model misread a bar chart, you'll understand why this matters.
A note on the numbers: the Sonnet 5 Terminal-Bench figure (10.3%) is what Anthropic's page reports. It's an unusually big jump, so for your own use case, run your own evals before you trust any single benchmark.
Claude Sonnet 5.5 Pricing
| Sonnet 5.5 | Opus 5.5 | |
|---|---|---|
| Input (per 1M tokens) | $2 | $4 |
| Output (per 1M tokens) | $10 | $20 |
| Cache writes (per 1M tokens) | $2.50 | $5 |
| Cache reads (per 1M tokens) | $0.20 | $0.20 |
Sonnet 5.5 costs half as much as Opus 5.5 for input, output, and cache writes. Add the per-task efficiency gains and the savings for high-volume workloads can be large.
Teams with strict data requirements can also get zero data retention.
What Early Customers Are Saying
The benchmarks look good, but results from real companies are more convincing. A few examples Anthropic shared:
- Slack: "Without changing prompts," Sonnet 5.5 beat Sonnet 5 on nearly all of their evals while using fewer steps and about 14% fewer output tokens.
- Zendesk: It resolved support tickets faster than the Claude models they currently run in production, working through cases about 20% quicker.
- Base44: Across 118 real app builds, it scored level with Opus 5 but needed an average of 3.6 iterations instead of 7.7. That's less than half the back-and-forth.
- Unity: It completed 90% of tasks in their multi-step benchmark.
A pattern shows up across these: fewer steps, fewer tokens, fewer retries. That's where the "up to 30% cheaper per task" claim comes from.
New Features and Improvements Worth Knowing
Better at long, agentic work
Sonnet 5.5 holds up better on long-running tasks and handles tool calls more efficiently, including batching them. For agent builders, that means fewer wasted round-trips.
Adjustable effort levels
You can set how hard the model thinks: Low, Medium, High, Xhigh, or Max. Use Low for quick classification and Max for tough debugging, and only pay for as much reasoning as the task needs.
Structured outputs
Structured outputs are supported, which makes it much easier to get reliable JSON into your application.
Sharper design sense
Anthropic says the model is better at design and polishing UI, which will matter to anyone using it to build front-end components.
The fun one: it beat Pokémon Red
Sonnet 5.5 is "the first Sonnet model to beat Pokémon Red working only from screenshots." It sounds like a gimmick, but finishing a long game from pixels alone takes a lot of planning, memory, and visual understanding over many hours.
Safety Updates
Anthropic shipped several safeguards with this release:
- Cybersecurity: safeguards comparable to Opus 5.5. Higher-risk requests fall back to Sonnet 5, and verified security professionals can apply for broader access through an expanded Cyber Verification Program.
- Biology: the same safeguards as Sonnet 5. They target harmful requests while keeping most research, education, and clinical work possible.
- Distillation protection: this is the first Sonnet model to launch with classifiers that stop others from extracting its reasoning to train copycat models.
Migrating from Sonnet 5 to Sonnet 5.5
For most teams, switching is a one-line change to the model ID:
import anthropic
client = anthropic.Anthropic()
response = client.messages.create(
model="claude-sonnet-5-5",
max_tokens=1024,
messages=[{"role": "user", "content": "Refactor this function for readability..."}],
)
print(response.content[0].text)
One gotcha: if you had thinking disabled on Sonnet 5, Anthropic says you'll need to switch the thinking setting to between_tools when you migrate. Check the official migration guide in the Claude Platform docs before you push to production.
Should You Switch to Claude Sonnet 5.5?
Switch now if you:
- Run coding agents or CI automation (it's the Terminal-Bench leader)
- Handle high-volume work like customer support, data extraction, or document processing
- Build computer-use or browser automation agents
- Currently use Opus for tasks that don't need it
Stick with Opus 5.5 if you:
- Work on very hard, novel engineering problems (Opus leads FrontierCode by about 8 points)
- Need every last bit of reasoning quality and cost isn't a concern
The practical approach: make Sonnet 5.5 your default and send only the hardest tasks to Opus. Many teams will see better results at a lower cost that way.
Final Thoughts
Model releases usually come with a trade-off: faster but less accurate, or smarter but pricier. Claude Sonnet 5.5 is unusual because it gets faster, cheaper per task, and more capable at the same time.
It doesn't replace Opus 5.5, and Anthropic doesn't claim it does. But for most day-to-day work, it's now hard to justify paying twice as much.
Try it on your own workload, measure tokens per task, and compare. The numbers may change which model you use by default.
Frequently Asked Questions
When was Claude Sonnet 5.5 released? Anthropic released Claude Sonnet 5.5 on September 28, 2026.
How much does Claude Sonnet 5.5 cost? $2 per million input tokens and $10 per million output tokens, half the price of Claude Opus 5.5.
Is Claude Sonnet 5.5 better than Opus 5.5? On agentic coding (Terminal-Bench 4.0) it scores higher: 70.6% vs. 66.4%. On most other benchmarks Opus 5.5 leads, but usually by only a small margin.
How much faster is Claude Sonnet 5.5? It generates output more than 30% faster than Sonnet 5.
Where can I use Claude Sonnet 5.5? In the Claude apps, on the Claude Platform (API), and through Amazon Web Services, Google Cloud, and Microsoft Azure.
What is the model ID for Claude Sonnet 5.5?
claude-sonnet-5-5
Source: Anthropic – Claude Sonnet 5.5