Reasoning models: when paying for thinking pays off

What "thinking" means, in plain terms

A reasoning model writes intermediate steps before its final answer: it breaks the problem down, tries an approach, checks it and sometimes starts again. The research term is test-time compute: rather than a bigger model, you spend more computation on each question.

Those steps are tokens. You usually see a summary or nothing, but you pay for all of them at the output rate, and they take time to generate. More thinking helps most on problems with many dependent steps, and little when the answer is a lookup or a label.

The packaging changed. In December 2024 OpenAI sold reasoning as a separate model, o1, at $15/$60 per million tokens; in 2026 it is a dial on the main models and OpenAI is retiring the o-series: o1 and o4-mini shut down on October 23, 2026, and o3 shuts down on December 11. On Claude Opus 5.5 and Fable 5.1, thinking cannot be switched off, so effort is the only lever.

How vendors expose it in October 2026

Same idea, different names, levels and defaults.

Three launches changed the table since late September: Haiku 5.5 (October 7) is the first Haiku with effort, Gemini 4 Argon (generally available since October 13) uses the same three levels as Flash, and GPT-6.1 Terra (October 26) accepts "none" again. For GPT-6 Astra, the model page lists low to max without stating a default, and OpenAI's reasoning guide states that "none" returns an HTTP 400 error.

The same level does not mean the same thing on two models. Anthropic says Opus 5.5 thinks more per turn than Opus 5 at the same level, that Opus 5.5 at medium beats Opus 5 at high on its coding and knowledge-work evaluations, and that Sonnet 5.5's levels are recalibrated against Sonnet 5. Effort also changes what the model does: at lower effort, Claude makes fewer and terser tool calls.

The cost and latency mechanics

Thinking is billed as output, the expensive side of the price list.

Anthropic, OpenAI and Google bill thinking tokens at the output rate, five times the input rate on Claude Opus 5.5 ($4/$20), GPT-6.1 Sol ($2/$10) and Gemini 4 Argon ($2/$10 until November 30, 2026, $4/$20 from December 1). Output is generated token by token, so thinking also delays the first useful word, and in an agent that delay repeats at every turn.

ARC Prize publishes score and cost per task for each model at each effort level, so the trade-off can be read straight from its data.

From low to high, Opus 5.5 gained 23 points for 1.7 times the cost; from high to max, cost rose 4.5 times and the score fell 1.6 points. ARC puzzles are not your workload, but that shape (a steep gain, then a plateau at rising cost) is what to look for in your own data.

GPT-6 Astra draws the same plateau on ARC Prize (85.4% at low, 92.1% at medium and at high, 93.3% at xhigh, 95.0% at max), as does GPT-6.1 Sol (76.7% at low, 91.7% at high and at xhigh, 94.2% at max). The Artificial Analysis index, which moved to version 5.0 in late October with nine evaluations of knowledge work, coding, science and general capability, shows the same shape:

From low to high, Opus 5.5 gains 17 points for 3.5 times the cost; from high to max, 6 points while cost almost triples. Astra flattens sooner: from high to max it adds 2 points for 1.9 times the cost. Astra scores higher up to high and costs less per task from high upward because it uses far fewer tokens (at max, 54 million against 220 million for Opus 5.5), even though each token costs 2.5 times more.

Vendors describe the same pattern. Anthropic reported in November 2025 that Claude Opus 4.5 at medium effort matched Sonnet 4.5's best SWE-bench Verified score with 76% fewer output tokens; OpenAI says GPT-6.1 Terra at max is more than three times as token-efficient as Claude Opus 5.5 on Agents' Last Exam, a computer-use test, at roughly a ninth of the cost per task. To measure that relation between effort, cost and score on your own workload, Slash offers AI consulting.

Where thinking pays, and where it doesn't

Thinking pays when a task has many dependent steps and a checkable result: math, code changes across files, planning, analysis that combines tools and data, and multi-step agents where one early mistake ruins the run. There, more effort raises the share of tasks solved, which is what you are buying.

It rarely pays for classification, extraction, routing, short FAQ answers or rewriting, where the model knows the answer or doesn't. It hurts real-time voice, where hidden thinking delays the first spoken word; Google offers Gemini 3.8 Live and a separate Live Extended Thinking variant, so you pick one or the other by the latency your use case can accept.

How to tune effort in five steps

1. Build an eval set. 30 to 50 real tasks per route, with automatic pass/fail checks; Anthropic recommends starting with 20 to 50 tasks drawn from real failures (see our evals guide).
2. Sweep the levels. Run each candidate at two or three effort levels, several times per task, recording success, tokens and latency.
3. Compute cost per solved task. Divide spend by tasks solved, not by requests: a cheap setting that fails half the time can cost more per solved task.
4. Set effort per route, explicitly. Keep the lowest level within your quality tolerance and write it in code, because defaults change (high on Opus 5, medium on Opus 5.5).
5. Monitor. Log reasoning tokens and stop reasons, alert when cost per task drifts, and re-run the sweep at every model change.

The output cap includes thinking: with a tight max_tokens (Claude) or max_output_tokens (OpenAI, Gemini), the model can spend the budget thinking and return a cut or empty answer, so leave headroom (OpenAI recommends reserving at least 25,000 tokens when you start) and treat a max_tokens stop reason or an incomplete response as an error. On Claude, changing the top-level effort between requests invalidates the prompt cache (per-message effort, in beta on Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1, preserves it), so keep effort stable within a cached conversation.

Pitfalls to avoid

• Overthinking lengthens turns, adds tool calls and can lower the score. On ARC-AGI-2, Opus 5.5 at max scored 1.6 points below high for 4.5 times the cost. Anthropic also notes that on Opus 5 effort does not reliably shorten visible answers, so ask for length in the prompt; OpenAI offers a separate verbosity setting.
• Settings copied from an earlier model can break. On Opus 5.5, thinking can no longer be disabled, forced tool use returns an error and the default effort fell from high to medium; on Sonnet 5.5, "disabled" returns an error and is replaced by "between_tools"; Google deprecated temperature, top_p and top_k in the Gemini API on July 21, 2026. Re-run the sweep instead of copying parameters.
• Asking Opus 5.5 to reveal its internal reasoning does not work. The model adds a reasoning_extraction refusal category that blocks attempts to extract its internal reasoning (returned as stop_reason: refusal). For an audit trail, ask for a short justification in the answer, log tool calls and route refusals to a fallback model.
• Written reasoning is not a faithful log of what the model did. OpenAI says GPT-6 Astra's written reasoning was harder to monitor than GPT-5.6 Sol's in its tests: Astra is better at controlling what it writes in its chain of thought (60.9% on the system card's controllability test, against 16.1% for GPT-5.6 Sol), and the UK AISI found reasoning summaries missing for up to 80% of long cyber trajectories. After the Hugging Face incident OpenAI committed more compute to chain-of-thought monitoring. Audit actions, tool calls and outputs.
• A cost dashboard with only input and visible output hides the part of the bill that effort controls. Log reasoning tokens as a separate line.

What it means for budgets in Colombia and Europe

Vendors bill in US dollars, so we budget reasoning in USD and convert at invoice time; for Colombian teams the exchange rate is one more variable outside their control. The unit that matters is cost per solved task, per route.

An illustration with the ARC Prize figures above, not a forecast: at 10,000 tasks a month, Opus 5.5 would cost about $2,410 at low effort, $4,080 at high and $18,500 at max, and on that test the step to max buys nothing.

Three levers keep thinking affordable: batch APIs (50% off at Anthropic, OpenAI and Google) for high-effort work that can wait, caching for the context agents resend at every turn, and routing that sends only hard cases to high effort. EU teams that need regional processing pay 10% more with OpenAI or on Claude's regional cloud endpoints, on a bill thinking has already raised. More in our AI cost guide; Slash AI Lab evaluates models before every deployment.