For most coding and agent work, Sonnet 5.5 is now the default I’d reach for, and Opus 5.5 is the exception you justify. At list price Sonnet costs exactly half of Opus in every workload I costed, and the small independent tests published so far found it cheaper per finished task too. The one thing I’d warn you about: at maximum effort Sonnet can think for so long that it stops being cheap, so set the effort level on purpose. And if you run agents, turn on prompt caching before you compare models at all, because caching changed the bill more than switching models did.
Live API prices: to compare what Sonnet 5.5 and more than 400 other models cost per million tokens today, see the AI models table on our dashboard, read twice a day from OpenRouter’s public model list.
On my own Pro plan I mostly pick Opus, not Sonnet, so I can’t compare Sonnet 5.5 with Sonnet 5 from daily use. The numbers below come from Anthropic’s price list and my own arithmetic.
Sonnet 5.5 at a Glance
| Released | 28 September 2026, six days after Opus 5.5 |
|---|---|
| API model ID | claude-sonnet-5-5 |
| Context and output | 1M-token context window, 128K-token maximum output |
| Knowledge cutoff | June 2026 |
| Where to use it | Claude apps (including the free plan, with its limits), Claude Code, the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry |
| Anthropic’s claim | About 30% faster and up to 30% cheaper than Sonnet 5 on typical workloads |
What It Costs
| Per million tokens | Sonnet 5.5 | Opus 5.5 |
|---|---|---|
| Input | $2 | $4 |
| Output | $10 | $20 |
| Cache write, 5 minutes | $2.50 | $5 |
| Cache write, 1 hour | $4 | $8 |
| Cache read | $0.10 (cut from $0.20 on 7 October) | $0.20 |
| Batch input / output | $1 / $5 | $2 / $10 |
Checked against Anthropic’s pricing page on 15 October 2026. Two details worth knowing: Sonnet 5’s $2 and $10 rate, originally introductory, never rose to the planned $3 and $15, and Sonnet 5.5 inherited it. And there’s no long-prompt surcharge on Sonnet 5.5; the full 1M context is billed at standard rates.
Our Calculation: Three Workloads at Today’s Prices
Method: we put Anthropic’s current list prices (checked 15 October 2026) into a short script and costed three fixed workloads. This is arithmetic, not a benchmark. It assumes both models use the same tokens and succeed first time, which the tests below show is often not the case.
| Workload | Sonnet 5.5 | Opus 5.5 | Haiku 5.5 |
|---|---|---|---|
| 100,000 short requests (1,500 in / 150 out) | $450 | $900 | $22.50 |
| 1,000 agent sessions × 20 turns, with caching (36K cached + 4K new in, 800 out per turn) | $478.40 | $956.80 | $27.34 |
| The same sessions without caching | $1,760 | $3,520 | $88 |
| Batch API, 100,000 short requests | $225 | $450 | $11.25 |
At list price, Sonnet is always half of Opus in these scenarios, so the real question is how often Opus saves you a retry or a human fix. And an agent without caching cost almost four times as much as the same agent with it. Fix caching before you argue about models.
The cache-read cut changed one popular comparison. Web Reactiva’s launch example (20 million cached tokens read, 1 million uncached input, 1 million output) came to $16 for Sonnet and $28 for Opus at launch prices, a 57% ratio. Recalculated with today’s prices, it’s $14 against $28: exactly half again.
Where Per-Token Price Misleads
- Effort changes the bill. Artificial Analysis, as reported by Web Reactiva, scored Sonnet 5.5 at maximum effort at 56 on its Intelligence Index, against 58 for Opus 5.5. To get there, Sonnet used about 193,000 output tokens per task, roughly $7.60 per task and about 50% more than Sonnet 5. Some analyses say Opus 5.5 at extra-high effort reaches the same score for less. Anthropic’s “up to 30% cheaper” claim refers to default effort, so both can be true.
- Defaults differ. In the API, Sonnet 5.5 starts at high effort and Opus 5.5 at medium; in Claude Code, Sonnet 5.5 starts at medium. Anthropic’s own table shows CursorBench scores of 39.2% at medium, 47.8% at high, 53.1% at extra-high and 55.5% at max for Sonnet 5.5. Set the same effort on both models before comparing them.
What Changed Beyond Price
Coding and agent scores. Anthropic’s launch table, as reproduced by Web Reactiva and Tododeia, shows 70.6% on Terminal-Bench 4.0 against 10.3% for Sonnet 5 and 66.4% for Opus 5.5. Opus still leads on FrontierCode, which measures whether code can be merged without human fixes. One commentator notes that part of the jump over Sonnet 5 reflects a newer, harder test, so don’t read it as seven times more capable.
Safeguards. Sonnet 5.5 is the first Sonnet with Opus 5.5’s cybersecurity safeguards. In the Claude apps and Claude Code, requests flagged as high-risk cybersecurity are re-run on Sonnet 5 with a notice. Ordinary development shouldn’t notice; security and low-level work deserve a test first.
Customer claims in the launch material, such as about 20% faster ticket handling and about 14% fewer output tokens, are vendor-selected testimonials, not measurements.
Early Independent Tests
Small tests published by others. I haven’t repeated them; I read Web Reactiva’s article in full and saw excerpts of the other two.
- CodeRabbit had Sonnet 5.5 review code with 13 known hard problems. It found 6, against 8 for standard Opus 5.5 and 10 for Opus at max, in about half the time and at roughly 60% lower cost. CodeRabbit planned to move simple and medium reviews to Sonnet and keep hard ones for Opus.
- A writer at The New Stack ran each model five times on three coding tasks against hidden tests. Sonnet 5.5 was perfect on all 15 runs and Opus 5.5 on 13, with Sonnet costing $12.69 against $22.07, about 42% less. Sonnet needed a higher output limit because it thinks longer per step.
- A 48-run Claude Code test published by wmedia.es found Sonnet 5.5 cheaper and faster than Opus 5.5 and Sonnet 5 on all four tasks, at about $0.88 for 12 runs against $1.91 for Opus. Small sample.
CodeRabbit’s split is the pattern I’d copy: Sonnet for the bulk of the work, Opus where a missed problem is expensive.
Check Before You Migrate
According to Web Reactiva’s reading of Anthropic’s documentation, effort levels were recalibrated, so medium on Sonnet 5.5 isn’t the same amount of reasoning as medium on Sonnet 5. Code that previously disabled thinking needs a different setting that keeps thinking between tool calls, which works at low, medium and high effort only. Review your requests before switching the model ID, and keep Sonnet 5 as a fallback until you’ve measured error rates on real traffic.
How to Decide
| If your work is… | I’d use |
|---|---|
| Everyday coding, agents and knowledge work | Sonnet 5.5 at medium or high effort |
| Code review where a missed bug is costly | Opus 5.5 for the hard reviews, Sonnet for the rest |
| High-volume, simple steps | Haiku 5.5 |
| Long agent sessions with repeated context | Sonnet 5.5 with caching on |
| Security or low-level systems work | Test first, because of the new cyber safeguards |
The number that should decide it is cost per accepted result, not cost per token. Run identical prompts on representative inputs, count accepted results, retries and cache hits, and price them at the current rate card.
FAQ
Is Claude Sonnet 5.5 free?
It’s the model on claude.ai’s free plan, with that plan’s usage limits, as Simon Willison noted at launch. API usage is billed separately.
Does a lower token price guarantee a lower bill?
No. A workflow can use more tokens, call more tools or need retries. Compare cost per completed task.
Is Sonnet 5.5 better than Opus 5.5?
Not in every case. It’s better value for most work; Opus still leads on the hardest problems and on code that has to merge without fixes.
Related: A 30-day Sonnet 5.5 learning plan · Claude Haiku 5.5 explained · AI tool pricing compared
1 comment