Anthropic released Claude Sonnet 5.5 on 28 September, and the price has not moved: $2 per million input tokens and $10 per million output tokens, the same as Sonnet 5. The benchmarks did move. On GDPval-AA, a knowledge-work benchmark in Anthropic's launch table, Sonnet 5.5 scores 1844 against 1846 for Claude Opus 5.5, which costs twice as much per token. If you run a SaaS product on Claude, your default model choice probably needs another look this week.
What Sonnet 5.5 actually is
Anthropic describes Sonnet 5.5 as a faster, cheaper complement to Opus 5.5. It is meant for well-scoped everyday work such as fixing bugs, drafting documents and running agent loops. Opus stays the pick for harder, open-ended problems. The model ID is claude-sonnet-5-5 on the Claude API, Google Cloud and Microsoft Foundry, and anthropic.claude-sonnet-5-5 on Amazon Bedrock. It is also available in the Claude apps.
According to Anthropic's docs, it has a 1M-token context window, up to 128K output tokens (300K on the Batch API with a beta header) and a June 2026 knowledge cutoff. The default effort level is high. For comparison, Opus 5.5 defaults to medium.
Anthropic says Sonnet 5.5 produces output more than 30% faster than Sonnet 5 and costs up to 30% less per task, mostly because it uses fewer tokens and tool calls to finish the same job. That is the vendor's claim. Customer reports quoted at launch point the same way: Slack reported 14% fewer output tokens, and Lovable reported about a third fewer tool calls.
The price, in full
From Anthropic's pricing page:
Standard: $2 input, $10 output. Cache reads: $0.20. Cache writes: $2.50 (5-minute) or $4 (1-hour). Batch API: $1 input, $5 output.
The full 1M context is billed at the standard rate, with no long-context surcharge. One detail is easy to miss. Sonnet 5's $2/$10 was originally introductory pricing, and a rise to $3/$15 had been scheduled for 1 September. Anthropic's pricing page now says that increase will not happen, so $2/$10 is the standard Sonnet price.
That puts Sonnet 5.5 at the same list price as GPT-6 Sol ($2/$10, per VentureBeat) and at half the price of Opus 5.5 ($4/$20).
How it compares
These are Anthropic's launch figures. Read them as marketing until your own evals agree:
Terminal-Bench 4.0 (agentic coding): Sonnet 5.5 70.6%, Opus 5.5 66.4% at its highest effort, Sonnet 5 10.3%. This is the one place in the table where Sonnet beats Opus.
CursorBench 4.0: Sonnet 5.5 55.5%, Opus 5.5 57.8%, Sonnet 5 34.1%.
OSWorld 2.1 (computer use): Sonnet 5.5 80.1%, Opus 5.5 81.8%, Sonnet 5 57.0%.
FrontierCode 1.1: Sonnet 5.5 46.2% at max effort, Opus 5.5 54.4%, GPT-6 Sol 49.3%. On this harder test Opus keeps a clear lead, and Anthropic's own table puts Sol ahead of Sonnet.
GDPval-AA v2.1: Sonnet 5.5 1844, Opus 5.5 1846, GPT-6 Sol 1487. Anthropic footnotes the Sol figures, so treat that comparison with care.
In short, Sonnet 5.5 gets close to Opus on routine agentic and office work and falls behind on the hardest coding. Anthropic says as much itself.
What Hacker News made of it
The launch reached the Hacker News front page with more than 750 points. Several commenters asked what Sonnet is for if Opus 5.5 at low effort is already as capable and as cheap in practice. Some reported good results on small builds. Others repeated a familiar complaint: in code review, models still weigh a typo and a security hole the same way. These are individual impressions, not measurements, but the Opus-at-low-effort question is worth testing yourself.
Migrating from Sonnet 5: read this first
This is not a drop-in model-string swap. Anthropic's docs list several breaking changes for code already on Sonnet 5:
Forced tool use now returns an error. Thinking blocks are tied to the model and conversation that produced them. The older computer_20251124 tool is not accepted on the Claude API or Google Cloud. The advisor tool rejects Sonnet 5 and older Opus models as advisors. Up-front thinking is now turned off with a new between_tools setting. Setting a non-default temperature, top_p or top_k returns a 400 error.
The one most likely to catch out a SaaS product: text between tool calls now comes back in thinking blocks. If your UI streams that text to users, it will go quiet between tool calls until you set a display value. Test your streaming chat before you switch production traffic.
Anthropic also says higher-risk cybersecurity requests will visibly fall back to Sonnet 5. If you build security tooling, check how that shows up in your responses.
What a small SaaS team should do
1. Make Sonnet 5.5 your default and route up to Opus on purpose. Send support replies, summaries, extraction and routine coding to Sonnet. Send only the long, ambiguous tasks your evals show Sonnet failing to Opus 5.5. OpenRouter, LiteLLM or a simple if-statement in your backend is enough.
2. Do the cost maths on your own traffic. Take an illustrative month of 10,000 requests, each with 3,000 input tokens and 800 output tokens. That is 30M input and 8M output tokens, which costs $140 on Sonnet 5.5 and $280 on Opus 5.5 at list prices. Put the non-urgent jobs through the Batch API and that Sonnet bill halves. Cache long system prompts too: reads cost $0.20 per million tokens.
3. Lower the effort setting. Sonnet 5.5 defaults to high effort. For classification or short replies, try medium or low and compare quality. VentureBeat, working from the vendor's figures, reported large per-task savings at lower effort.
4. Run a small eval before switching. Pull 50 to 100 real prompts from your logs and run them through Sonnet 5, Sonnet 5.5, Opus 5.5 at low effort and GPT-6 Sol. Promptfoo or Braintrust will do the job. Compare quality, latency and actual cost per task. Tokens per task matters more than price per token.
5. Leave room for Haiku. Anthropic says Haiku 5.5 is due "in the coming weeks". For high-volume, low-stakes calls, it may be worth waiting for.
The bottom line
Sonnet 5.5 costs the same as before and, on Anthropic's numbers, comes close to Opus 5.5 on everyday agentic work. That makes it a sensible default for most SaaS workloads at half the Opus token price. Plan for the breaking changes, especially the streaming one. Test it on your own prompts against Opus at low effort and GPT-6 Sol, and switch only once the cost per task is lower in your logs, not just on the launch page.