Claude Sonnet 5 costs 40 per cent of Opus and counts tokens differently. What to re-check before switching

Update, 11 August 2026: Anthropic has made Sonnet 5's introductory pricing of $2 per million input tokens and $10 per million output permanent. The planned move to $3 and $15 on 1 September no longer applies. The rest of this piece stands.

Anthropic released Claude Sonnet 5 on 30 June and called it the most agentic Sonnet yet. The claim that matters to a buyer is in the pricing paragraph: $2 per million input tokens and $10 per million output, against $5 and $25 for Opus 4.8, for a model Anthropic says performs close to Opus 4.8 at higher effort settings. Sonnet 5 is now the default model on Claude's Free and Pro plans, and it is available on the Claude API, on AWS, on Google Vertex and in Microsoft Foundry from day one.

For most agent workloads, the mid-tier model is now the sensible default rather than the compromise. That is a change in how you should choose models, and it comes with three details that decide whether the saving on paper turns up on the invoice.

The tokeniser changed, so your token counts did too

Sonnet 5 uses an updated tokeniser, the same kind of change Anthropic made with Opus 4.7. Anthropic's own footnote says the same input can map to roughly 1.0 to 1.35 times as many tokens as before, depending on the content.

That means a workload that used to cost 1,000 tokens of input on Sonnet 4.6 may cost anywhere from 1,000 to 1,350 on Sonnet 5, and the multiplier is different for prose, code, tables and non-English text. At the introductory price the saving still holds at the top of that range. At the standard price of $3 and $15 that was scheduled for September, a workload at the 1.35 end would have cost about a third more than it did on Sonnet 4.6, at the same list price.

The practical instruction is short. Do not estimate the new cost by multiplying your old token counts by the new price. Run a representative sample of your real inputs through the new model, read the token counts off the response, and use those.

Effort is now a dial, and the default is not always right

Sonnet 5 exposes effort levels, and Anthropic's own cost-performance charts show why they matter. At medium effort the model is much cheaper than Opus 4.8 for a given score. At the highest setting it can match Opus 4.8 on some agentic tasks, and it spends the tokens to do so. Anthropic raised rate limits across its products at launch specifically to accommodate the higher token use of higher effort.

So "switch to Sonnet 5" is three decisions, one per effort level, and the right one varies by task. Extraction and classification steps will usually pass your tests at low effort. Planning steps in an agent loop may need high effort to match what Opus was doing. The job is to run your evaluation set at each level and pick the cheapest setting that passes, per step, rather than picking one setting for the whole system.

Parameters that no longer exist

Anthropic has deprecated temperature, top_p and top_k on Sonnet 5, as it did on Opus 4.7 and later. Requests that set them to a non-default value get an error rather than a warning.

This is a small thing that stops production. Most codebases that have been calling Claude for more than a year set a temperature somewhere, often in a shared client wrapper nobody has opened since 2024. Search for the three parameter names before you change the model string, because the first failed request will otherwise be in production.

What actually improved, and what did not

Anthropic's evaluations describe Sonnet 5 as a strict improvement over Sonnet 4.6 on reasoning, tool use, coding and knowledge work, with lower rates of hallucination and sycophancy and better resistance to prompt injection in agentic settings. Early access partners report the thing that matters most in practice: it finishes multi-step tasks that earlier Sonnet models abandoned halfway, and it checks its own output without being asked.

Two caveats from the same system card are worth keeping in view. On Anthropic's behavioural audit, Sonnet 5 shows a lower rate of misaligned behaviour than Sonnet 4.6 but a higher rate than Opus 4.8. And it was not trained for cybersecurity work, so teams doing security research should stay on Opus.

Neither caveat changes the default for ordinary business agents. Both are reasons to keep the same human checkpoints you had before, because a cheaper model that finishes more tasks is also a model that does more things unsupervised.

The four checks, in order

Search the codebase for the three retired parameters and remove them.

Take 200 real inputs per workload, run them through Sonnet 5, and record the token counts. That is your new cost basis.

Run your evaluation set at low, medium and high effort for each step of each workload, and record the pass rate and the cost at each. Pick the cheapest passing setting per step.

Shadow-run for a week against the model you are replacing, diff the outputs, and read the disagreements before cutting over. The migration playbook we published last month covers the rest.

If you do only one of these, do the second. The tokeniser change is the one that turns a 60 per cent saving into a 40 per cent saving without anyone noticing, and the invoice at the end of the month is the wrong place to find out.

What this means for the market

The gap between the mid-tier and the top tier has been closing for a year, and Sonnet 5 closes it further. When a model at two-fifths of the price does most of what the flagship does, the flagship becomes the tool for the hard step in the loop rather than the default for every call. That is good for buyers, and it is the reason we recommend clients build per-step model selection into their agents rather than hard-coding one model everywhere. It is also the structure our AI products team ships by default, because the next release will move the line again.

Frequently asked

How much does Claude Sonnet 5 cost? $2 per million input tokens and $10 per million output tokens, which Anthropic launched as introductory pricing on 30 June 2026 and made permanent on 10 August 2026. Opus 4.8 is priced at $5 input and $25 output. Sonnet 5 uses an updated tokeniser that can map the same text to roughly 1.0 to 1.35 times as many tokens as Sonnet 4.6, so measured cost per task should be checked rather than assumed.

What changed in Claude Sonnet 5 for developers? Adjustable effort levels that trade cost against capability, a new tokeniser, and the removal of the temperature, top_p and top_k parameters, which now return an error when set. Anthropic reports better tool use, coding and multi-step task completion than Sonnet 4.6, with lower hallucination and stronger resistance to prompt injection, and weaker cybersecurity capability than Opus.

Should you switch from Opus to Sonnet 5? For most business agent workloads, probably yes, but per step rather than wholesale. Run your evaluation set at each effort level for each step, choose the cheapest setting that passes, and keep Opus for the steps that fail on Sonnet. Measure real token counts on your own inputs first, because the tokeniser change affects different content differently.

Related reading

• Your vendor will retire the model you built on
• Flat-rate AI is ending for agents
• Building evaluation harnesses for production AI systems

A cheaper model is only cheaper once you have measured it on your own work. Sonnet 5 is a good reason to build the measurement, and the measurement will outlast the model.