Cost per task: Sonnet 5.5 costs the same per token and 50% more per task

A launch post says the new model costs less per task; an independent measurement says it costs more. Anthropic claims Sonnet 5.5 can cost up to 30% less than its predecessor. Artificial Analysis’ launch-day analysis on September 28 put it roughly 50% above, on its own evaluation mix. The missing line on the price list is how much work the model generates before it finishes your task.

Cost per task is the bill, not the rate

Per-token prices describe what you buy. Cost per task describes how much you consume to get an outcome. A lower rate cannot guarantee a lower bill if the model generates more reasoning, repeats failed steps, or needs another attempt.

Artificial Analysis defines its measure as a weighted average cost per Intelligence Index task, including input, cache hits, cache writes, reasoning, and answer tokens. Its Index v4.3.2 combines ten evaluations: Agents 30%, Coding 20%, Scientific Reasoning 20%, and General 30%. Those weights describe its workload, not yours.

The authors explicitly warn that the metric “has limitations and may not apply directly to every use case.” For your operation, define a completed unit of work and count failed attempts toward its cost. A cheap response that leaves the job unfinished is not a cheap completed task.

Sonnet 5.5: the same rate, different consumption

Anthropic lists Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, unchanged from Sonnet 5. It says reduced token consumption produces up to 30% lower cost per task in its testing. On several benchmarks, it also reports Low or Medium effort beating Sonnet 5’s highest score for about a tenth of the cost per task. These are vendor results under particular settings, not a blanket saving.

Artificial Analysis’ September 28 launch article reported $7.60 per task on its own Index mix, approximately 50% above Sonnet 5. That figure is already dated. The same organization’s model page, checked on October 20, lists $5.46 per Index task for the same configuration — a figure that moved by roughly a quarter in three weeks, in the direction that flatters the model. Hold that thought before you budget from anyone’s number.

At max effort, Artificial Analysis measured approximately 193k output tokens per Index task: around 60% more than Opus 5.5 or Sonnet 5 at max effort. Its explanation for the cost is consumption, not a higher headline output rate.

The sources also disagree on cache pricing. Anthropic’s page says cache reads cost $0.10 per million tokens, versus Sonnet 5’s $0.20. Artificial Analysis describes identical pricing at $0.2/$2/$10 per million cache-input/input/output tokens. Do not silently combine those claims. If caching materially affects your budget, confirm the current price sheet and test your actual cache usage.

This is not evidence that Anthropic’s saving is impossible. It is evidence that “less per task” needs a workload, effort setting, and measurement method attached.

Gemini 4 Argon’s measured cost depends on promotional pricing

Gemini 4 Argon illustrates a different mechanism. Artificial Analysis reports $1.99 per task on Artificial Analysis’s Index mix with an initial 50% pricing promotion and cache discounts increased to 95%. Without the promotion, its estimate rises to $3.98, above GPT-6 Astra at max effort at $3.26 on the same mix.

That saving does not come from shorter answers. Artificial Analysis measured Argon averaging 62k output tokens per task against Astra’s 27k. Lower token prices fund the current advantage.

That math has since resolved. Google announced global availability for Argon on October 13, 2026, through the paid Gemini API and specified subscription products, and set the introductory period to expire on November 30, 2026, after which standard $4 per million input and $20 per million output pricing applies. So the promotion is no longer an open question with no date attached — it has a deadline you can put in a calendar. Budget both price states, and know which one you are in. A discounted result is a conditional estimate, not a permanent property you can build a margin around.

A benchmark number belongs to its harness

A harness is the surrounding agent setup: instructions, tools, execution rules, and limits. Changing it can change both the score and the bill. Results from different harnesses do not establish a model ranking.

For Terminal-Bench 4.0, Anthropic reports Sonnet 5.5 at 70.6% in its evaluation setup. Artificial Analysis reports 64% in its mini-SWE-agent-style harness. Vals AI reports 64.14% under mini-SWE-agent. The latter results use similar methods; their close agreement is not two independent methodological confirmations.

Within Vals’ harness, Opus 5.5 scored 65.15% at $13.20 per test, versus Sonnet 5.5’s 64.14% at $16.51. That comparison is useful inside that setup. It does not license a ranking across different harnesses or turn those per-test charges into your operating costs.

Five moves before you change your default

1. Run a small comparison on your own queue. Pick representative tasks and decide what counts as an acceptable completion before you test anything. Run your current setup and one candidate against the same tasks. Record effort settings, total charges including failed attempts and retries, accepted completions, and elapsed time. Divide each setup’s total charges by its accepted completions. Public measurements choose your candidates; your own queue is what sets your budget.

2. Measure across effort settings. Sonnet 5.5 offers low, medium, high, xhigh, and max. Compare them against the same acceptance criteria, retries included, and include your current default effort setting in the comparison. Do not assume either higher effort or lower effort improves the total bill.

3. Split planning from execution where tasks allow it. Let a stronger model decide what to do, then send bounded execution to a cheaper, faster model that meets your requirements. Cascades and routers offer related allocation strategies, not proof that every workflow benefits from this exact split. Keep a check on the handoff and a route back when execution fails.

4. Test a specialist on one task type. Choose a recurring category where your default is expensive or unreliable, then compare a candidate against the default on that category before routing work to it. Count the cost of selecting the model, checking its output, and recovering from failures. Measure the whole routed workflow, not just the task you handed off.

5. Budget the undiscounted price. Record which rates are promotional and what the workflow costs without them. Argon’s estimate without the promotion shows why. If the workflow only works financially during a promotion, make that dependency explicit before committing volume.

Where the playbook breaks

Lower effort can create more work. Qwen’s Qwen3.8-27B model card warns that faster individual turns in multi-turn agent tasks can mean insufficient analysis, failures, and retries, increasing total latency and tokens.

More effort can also hurt. On FrontierCode, Anthropic reports that Sonnet 5.5 scored lower at Max than at Xhigh: in two cases Cognition examined, its code-review subagents led to a timeout or to out-of-scope edits. Spending more did not secure a higher score.

Quality requirements can disqualify a candidate before price matters. Artificial Analysis measured Sonnet 5.5 at 54% factual accuracy on AA-Omniscience, against Opus 5.5’s 66%. For factual-recall work, that gap is a reason to investigate task fit, not to excuse it with aggregate capability.

Measurement can lag deployment. Artificial Analysis tested a pre-release Sonnet build with a disclosed structured-output bug that Anthropic says is fixed publicly. A rerun was announced, but none had been found as of October 20, 2026. Argon’s launch was restricted to selected users at first, which limited operator replication during its first weeks.

Routing has its own boundary. IBM Research warns that predictive routers may be less accurate on queries unlike their training data. Novel work needs a fallback and review, not automatic faith in yesterday’s routing decision.

The research behind this approach

FrugalGPT, RouteLLM, and Hybrid LLM support testing whether routing between models can cut cost while holding quality steady. Their results depend on the models, workloads, and methods studied — the headline savings figures belong to 2023 and 2024 papers on GPT-4-era systems. They are not a forecast for your operation.

What this article does NOT cover

This is not a procurement comparison or a complete accounting of hosting, human review, security, and failure losses. It does not prescribe a router implementation. Your decision still needs a quality threshold and a definition of completed work that matches your operation.

Related guides

• Cost-aware model routing for agents
• Why every agent needs a cost cap on day one
• Claude Opus 5 guide

Sources

• Artificial Analysis Intelligence Benchmarking Methodology
• Anthropic: Claude Sonnet 5.5
• Artificial Analysis: Claude Sonnet 5.5
• Artificial Analysis: Gemini 4 Argon
• Artificial Analysis: Claude Sonnet 5.5 model page
• Google: Introducing Gemini 4 Argon
• Vals AI: Terminal-Bench 4.0
• Qwen3.8-27B model card
• IBM Research: LLM routers
• FrugalGPT
• LMSYS: RouteLLM
• Hybrid LLM