Choosing the right model
and effort for the job

Here's the figure that reframed this for me. On the Artificial Analysis leaderboard, read on 25 September 2026, Claude Opus 5.5 at medium effort scores 51 on their Intelligence Index for $1.34 per task. The same model at max effort scores 58 for $5.98. Same model, same price per token. Seven points cost four and a half times as much per task. GPT-6 Sol shows the same shape: 40 at medium for $0.25, 48 at max for $1.06.

Neither number is wrong to pay. Seven points can be worth it. But that's a decision about a specific job, and when effort is left at a default it doesn't get made.

Three dials, not one

Every AI workload in a company sits at the intersection of three choices. Which model tier. How hard it's told to think. How the tokens are bought. The first gets the attention, because it's the one on the price sheet. The second is now the largest lever on a single model. The third is close to free money for work that can wait or repeats itself.

Effort is a real dial now, at all three vendors

Anthropic's effort setting runs low, medium, high, xhigh and max, and it governs every output token: thinking, the visible answer, and tool calls. Lower effort also means fewer tool calls, which matters more than the thinking savings once an agent is running. Opus 5.5 defaults to medium, one level lower than Opus 5 did, and its thinking can't be switched off, so effort is the main thing standing between a request and its bill. Anthropic describes Sonnet 5 at medium as comparable to Sonnet 4.6 at high. At launch they claimed Opus 5.5 at its default beats Opus 5 at max on Terminal-Bench 4.0 "for about a fifth of the cost." That's a vendor claim about vendor benchmarks, so treat it as a hypothesis for your own evals.

OpenAI's reasoning effort works the same way. Reasoning tokens are never shown to you and are billed as output tokens. Google's Gemini thinking levels bill the same way, as output plus thinking tokens. Google's own launch post for Gemini 3.8 Flash says the new model "works harder" and may use more tokens, and it tells efficiency-first workloads they can stay on 3.7 Flash, which is still supported. That's a vendor telling you, in plain words, that the same price per token doesn't mean the same cost per task.

The price sheet is a floor, not a forecast

For the record, here's the September list price per million tokens, input then output. Opus 5.5 $4 and $20. Sonnet 5 $2 and $10, now permanent after a planned increase to $3 and $15 was cancelled. Haiku 4.5 $1 and $5. GPT-6 Astra $10 and $50, GPT-6 Sol $2 and $10, GPT-6 Luna $0.10 and $0.50. Gemini 3.8 Flash $0.75 and $3.75.

Then the adjustments the table doesn't show. Output costs five times input at Anthropic and OpenAI, and thinking is billed as output, so a model that thinks longer costs more than its input price suggests. Anthropic notes that Claude 4.7 and later use a tokenizer that produces about 30% more tokens for the same text, so a lower per-token price on a newer model doesn't translate one-for-one into a lower bill. Gemini 3.8 Flash's price is introductory and doubles on 1 January 2027. None of this is hidden. It's just not in the column people compare.

How you buy the tokens

Two levers need no model change at all. Batch APIs at all three vendors cost half the standard rate, for anything that can wait: overnight classification, back-catalog enrichment, eval runs. Prompt caching reads a repeated prefix at a tenth of the input price on most Claude models and a twentieth on Opus 5.5. A five-minute cache write costs 1.25 times the input price, so it pays for itself on the first reuse. The two stack. A large, stable system prompt fed through a batch job is the cheapest shape of work these platforms sell.

One caution from the same docs. On Claude models without per-message effort, changing effort mid-conversation starts the cache over. Pick the level at the start of a cached session, or use a model that can change it per message.

Route the work, don't average it

A company doesn't have one AI workload. It has a support queue, a document pipeline, an internal code assistant and a few people doing genuinely hard analysis, and one model at one effort level is the wrong answer for all of them at once. The published evidence for routing is getting concrete. RouteNLP, posted in April 2026, reports 40 to 85% lower cost across its benchmark tasks while keeping 96 to 100% of quality on structured tasks. Its eight-week enterprise pilot, at about 5,000 queries a day, cut inference cost 58% with 91% of responses accepted, and p99 latency fell from 1,847 ms to 387 ms. That's one paper and one pilot, not a law, but it matches the direction of everything else here.

My own build tooling routes by task type. Documentation and scaffolding go to the smallest tier. Real implementation goes to the middle tier. Cross-module design goes to the top. Anything that fails quality checks three times escalates to the top tier regardless of what it started as. The go-to-market run I wrote about this month, eighteen sections for thirty-nine cents, ran on the middle tier. The rule I hold hardest is the dull one: never spend the top tier on finding files.

The spend numbers say measure first

A DoiT survey of 500 finance leaders at US and UK companies with 1,000 or more employees, fielded in February 2026, found 79% had AI cost overruns in the prior twelve months. Only 15% could calculate AI return on investment without significant bottlenecks. The number I keep rereading: organizations that rated their cost management most mature reported more overruns, 89%, and bigger ones, 30.9% on average, than those just starting out, at 69% and 16.1%. I'd read that as the mature teams being the ones who can see their overruns. Either way, you can't tune a cost you're not measuring per job.

What I'd actually do

Set effort explicitly on every workload instead of inheriting a default, because the defaults changed this month and will change again. Run an effort sweep on your own evals before choosing a level. Anthropic's documentation says exactly this, and it's the only way to find out whether medium holds for your task. Measure cost per accepted task, not cost per token, because retries and rejected drafts are where the real spend hides. Route by task type, with an escalation path for the cases the cheaper tier gets wrong. Cache every stable prefix and batch everything that can wait a few hours. And redo the sweep on every model change, which is the same discipline I argued for in every model you depend on has a sunset date.

None of this needs a new platform. It needs someone to decide, job by job, how much intelligence each piece of work actually requires, then write that decision down where the next model change can't quietly undo it.

Sources, all read 25 September 2026: Artificial Analysis model leaderboard, cost-per-task column (artificialanalysis.ai/leaderboards/models; the index version isn't shown on the page, and the methodology was revised several times in early September). Anthropic pricing and effort documentation (platform.claude.com/docs, "Pricing" and "Effort"). Anthropic, "Claude Opus 5.5" launch post, 22 September 2026. OpenAI API pricing and reasoning guide (developers.openai.com). Google, Gemini 3.8 Flash launch post, 2 September 2026, and Gemini API thinking documentation. RouteNLP, arXiv 2604.23577, 26 April 2026. DoiT and Sapio Research, AI spending survey, published 9 June 2026. The routing table is from my own development tooling, not a benchmark. Prices change often, so check the vendor pages before relying on them.