Sonnet costs 36% less than Opus in coding test with retries, The New Stack says


The New Stack author's coding comparison finds Sonnet cheaper overall, including retries. Opus completes tool-enabled bug fixes faster and at lower cost.
Sonnet 5.5 and Opus 5.5 are Anthropic's artificial-intelligence models. The company launched Sonnet 5.5 six days after Opus 5.5.
A coding comparison published by The New Stack on 8 October found that Sonnet 5.5 cost about 36% less than Opus 5.5 overall, including four failed attempts—but Opus completed a tool-enabled bug-fixing task faster and more cheaply.
The author's final set of runs recorded 15 successes out of 15 for Sonnet and 13 out of 15 for Opus. Sonnet's tally included four successful reruns after it initially stopped at an output limit. The prompts were not published, so the results apply to the described workloads and settings rather than establishing which model is better generally.
Anthropic released Claude Sonnet 5.5 on 28 September. How2B reported on 29 September that Anthropic aimed Sonnet 5.5 at clearly defined everyday tasks, including fixing software bugs and preparing documents, slides and spreadsheets. Anthropic claimed output generation more than 30% faster than Sonnet 5 and task costs up to 30% lower through reduced token use. The company also reported improvements in clear writing, image understanding and design work. Those claims compare Sonnet with its predecessor, not with Opus.
The New Stack author ran each model five times on each of three coding tasks through Anthropic's API—the interface software uses to request model responses. Both models received identical prompts, used adaptive thinking to adjust their reasoning, and ran at the maximum effort setting. The author checked answers against tests the models could not see and recorded tokens, costs at listed prices and elapsed time.

| Workload or total | Sonnet 5.5 | Opus 5.5 |
|---|---|---|
| Tool-enabled bug fixing | 5/5 passed; 5 min 8 sec; 42,608 output tokens; 29 tool calls; $0.70 per run | 5/5 passed; 3 min 21 sec; 20,625 output tokens; 25 tool calls; $0.75 per run |
| Dependency resolver | 5/5 passed; 8 min 58 sec; 81,097 output tokens; $0.82 per run | 5/5 passed; 9 min 40 sec; 70,687 output tokens; $1.42 per run |
| Concurrency bugs | 5/5 passed; 12 min 13 sec; 101,788 output tokens; $1.02 per run | 3/5 passed; 16 min 43 sec; 111,428 output tokens; $2.24 per run |
| Successful runs in the final set | 15/15, including four successful reruns | 13/15 |
| Total time for the final 15 runs | 2 hr 11 min 36 sec | 2 hr 28 min 39 sec |
| Total cost, excluding Sonnet's initial failures | $12.69 | $22.07 |
| Total cost, including Sonnet's initial failures | $14.09 | $22.07 |
In this comparison, Opus completed the tool-enabled bug fixes about 35% faster. Sonnet completed the resolver slightly faster and at 42% lower cost, despite using about 15% more output tokens. On the concurrency task, Sonnet was faster and cost less than half as much. Across the final 15 runs, Sonnet finished about 11% faster overall; that timing does not include its initial failed attempts.
Tokens are the units used to count text a model processes and generates, including generated reasoning. Sonnet 5.5's API rates are unchanged from Sonnet 5 and are half Opus 5.5's rates. But a task's bill depends on how many tokens the model uses, not just the price of each token.
| Model | Input tokens | Output tokens |
|---|---|---|
| Sonnet 5.5 | $2 | $10 |
| Opus 5.5 | $4 | $20 |
Sonnet used about twice as many output tokens as Opus on the tool-enabled bug-fixing task, almost erasing its price advantage before retries. Four of Sonnet's first five attempts also reached the initial limit of 32,000 output tokens per step and stopped before finishing. Opus stayed well below the same limit. The author raised Sonnet's limit to 128,000 and reran those four attempts successfully.
The failed attempts added about $1.40. Including them raised Sonnet's average cost on that task to about $0.98 per run, compared with Opus's $0.75. Across all three workloads, the extra attempts reduced Sonnet's reported saving from 42% to about 36%. Sonnet's greater token use also explains why its half-price rates did not produce a 50% saving.
The New Stack also cited a different result from testing firm Artificial Analysis: at maximum effort, Sonnet cost $7.67 per task against Opus's $5.98 and scored slightly lower on its Intelligence Index. That comparison and the author's workloads produced different cost rankings. The author did not publish the prompts, explaining that the supporting code projects, specification and module were too long to include.
Analysis
Anthropic's launch figures include Terminal-Bench 4.0, which tests tasks requiring several steps. These are company-reported benchmark scores, not results from The New Stack author's coding comparison.
| Model | Score |
|---|---|
| Sonnet 5 | 10.3% |
| Sonnet 5.5 | 70.6% |
| Opus 5.5 at the xhigh effort setting | 66.4% |
On GDPval-AA, a test based on real work across different professions, Anthropic said Sonnet 5.5 nearly matched Opus 5.5. The company cautioned that individual tests do not capture all capabilities and said Opus retained an advantage in complex, sustained reasoning. How2B also reported early testers' assessments that Sonnet followed slide templates better and produced presentations needing less editing.
The release also added cybersecurity safeguards. For high-risk cybersecurity tasks, the system switches to the older Sonnet 5 model and notifies the user.
How2B reported on 29 September that Claude Sonnet 5.5 was also available through Amazon Web Services, Google Cloud and Microsoft Azure. The same report said Anthropic planned to release Claude Haiku 5.5 within weeks, targeting high-volume uses where cost matters most.
The Swedish firm SKF used SI to recreate the Hollywood star for a 99-second promotional film. It is the first time her likeness has been used in moving images since her death. One of the relatives who manage her legacy says he would not rule out more roles.
4 min

The smart-ring maker withdrew its planned Nasdaq share sale just over a week after filing for it. Nuclear firm Holtec International postponed its own listing earlier this month, blaming factors that it said had hurt investor confidence in new share offerings.
3 min

Adobe says its older CEP plugin system now depends on outdated parts. Developers of plugins for Photoshop, Premiere, After Effects, Illustrator, InDesign and Media Encoder now have dates for moving their tools to UXP, Adobe's newer system.
4 min