Claude Sonnet 5 is worth testing now if your team runs tool-using, coding or knowledge-work agents and can measure completion cost—not merely token price. It is not worth an immediate production switch because a launch benchmark or temporary discount looks attractive.
Anthropic says Sonnet 5 launched with introductory API pricing of $2 per million input tokens and $10 per million output tokens through 31 August 2026, before moving to $3 and $15 respectively. The deadline creates a useful evaluation window, but it can also create a false economy if a team optimises around a temporary price without modelling the standard rate.
The switching scorecard
| Workload | Test first | Switching signal | Stay signal |
|---|---|---|---|
| Coding agent | Repository task completion and regression rate | More tasks finish with fewer human repairs | Faster output creates more review work |
| Browser research | Source coverage, citation fidelity and recovery from blocked pages | Higher verified-answer rate per run | Confident synthesis outruns evidence |
| Document analysis | Extraction accuracy across long, messy files | Fewer missed clauses and stable structure | Costs rise because context is repeatedly resent |
| Tool workflow | Correct tool choice, argument accuracy and retry behaviour | Fewer failed or unnecessary calls | Agent loops or mutates when it should ask |
| Executive drafting | Accuracy, voice and decision usefulness | Less editing without unsupported claims | Polished prose hides factual errors |
Anthropic’s Claude Sonnet 5 launch post positions the model as a stronger agentic Sonnet for reasoning, tool use, coding and knowledge work. It also records a correction to an earlier BrowseComp chart methodology. That correction is a reason to read the current primary source and system card rather than copying a launch-day graphic into a purchasing decision.
Token price is not task price
A cheaper input token does not guarantee a cheaper completed task. Task cost includes:
repeated context and cache behaviour;
output length;
tool calls and external API fees;
retries after failed actions;
human review and repair time;
latency and rate limits;
the cost of an incorrect production action.
Calculate cost per accepted outcome. If Model A costs £0.40 in tokens but needs fifteen minutes of engineer repair, while Model B costs £0.90 and needs two minutes, the second route may be cheaper. If a high-effort setting consumes far more tokens without improving acceptance, the lower sticker price is misleading.
Run a 30-task shadow evaluation
Choose ten easy, ten representative and ten difficult tasks from real work. Remove private data unless the approved platform and contract allow it. Run the current production model and Sonnet 5 against the same task definition, tools and acceptance checks.
Score each run:
| Measure | Definition | Weight |
|---|---|---|
| Accepted outcome | Meets the task’s objective without material correction | 35% |
| Factual and source accuracy | Claims reconcile to provided or primary evidence | 20% |
| Tool discipline | Correct calls, no unsafe mutation, bounded retries | 15% |
| Human repair minutes | Time to make the output production-ready | 15% |
| End-to-end latency | Time from request to accepted result | 5% |
| Standard-price cost | Modelled at post-introductory rates | 10% |
Do not change the rubric after seeing which model wins. Keep failures, not only best examples.
Model both price periods
Create two cost columns: introductory price and standard price. A workload that is viable only until 31 August is a campaign, not a durable architecture. That can still be useful for a migration project or finite backlog, but label it honestly.
Also test effort settings. Anthropic says Sonnet 5 covers a range of cost-performance options and can approach higher-tier performance on some tasks at higher effort. The operative phrase is “some tasks”. Select effort by workload and evidence, not globally.
For teams routing different projects to different models, the GPT-5.6 workload router demonstrates the broader principle: tier choice belongs to the task, reversibility and evidence burden.
Migration risks that benchmarks do not show
Prompts can be model-specific. A workflow tuned around one tool schema, refusal style or context behaviour may regress when moved. Structured output can differ at edge cases. Safety behaviour can change. An agent that is more willing to act may require stricter permissions.
Before switching:
Pin the model identifier rather than relying on an unversioned default where possible.
Record current prompts, tool schemas and acceptance tests.
Run shadow traffic without production mutations.
Compare failures by category.
Set spending and tool-call limits.
Move a small reversible workload first.
Keep an immediate rollback route.
The Claude Sonnet 5 system card provides broader safety and evaluation context. It does not replace testing on your own data, tools and consequences.
A break-even calculation
For each workload, calculate monthly accepted tasks, tokens per task, retries and review minutes. Price the same volume at the introductory and standard rates. Then add an internal cost for human review. The exact labour figure can be approximate; using zero is the serious error.
If Sonnet 5 saves three review minutes on 10,000 monthly tasks, the operational gain may outweigh a higher token bill. If the model saves tokens but increases investigation after subtle tool mistakes, the apparent saving can disappear. Record both the mean and the worst ten failures because rare high-impact errors matter in agentic systems.
Do not annualise the launch discount. A procurement forecast should use the announced standard price from September, with the introductory period shown separately as a temporary benefit.
Who should wait
Wait if the workflow has no automated acceptance tests, if tool permissions are broad and unlogged, or if regulated data cannot enter the new service under the current contract. Also wait if the team cannot pin prompts, model versions and rollback conditions. In those cases, changing the model adds variance before the operating system can observe it.
The considered verdict
Use the introductory period to evaluate Sonnet 5 on a bounded, representative workload. Switch where accepted-task cost, evidence quality and human repair improve at both introductory and standard pricing. Keep the current model where the apparent gain disappears after review or where migration risk exceeds the benefit.
The deadline is useful because it creates a decision date. It should not become the reason for the decision.
