Which AI model is best for each job

Each job below has three picks: the strongest model, the best value, and the strongest open-weights model. Select a job to see the scores and the reasoning behind its picks.

Jobs no benchmark covers well yet

• Long-document analysis. Most models can now read about a million tokens in one request. That limit is called the context window. No current, independent benchmark measures how well a model uses all that text, so the table's context column shows capacity and says nothing about quality.
• Research with web search. LMArena's search rating has no entry for any model released after August. BrowseComp, a test of finding hard-to-locate facts on the web, has results for this month's models only from the model makers themselves, and I found no independent run. There is no pick until an independent measure catches up.

This month's take

Opinion, not measurement. Every number below comes from the tables on this page. The reading of them is mine.

1. Start at the $2 tier this month

   GPT-6.1 Sol and Claude Sonnet 5.5 cost $2 per million input tokens and score 166.1 and 165.0 on the overall index. Claude Opus 5.5 leads at 167.3, and all three sit inside each other's margins of error. The $10 tier buys nothing you can measure. GPT-6 Astra scores 166.4 and Claude Fable 5.1 scores 164.7, both below Opus 5.5, which costs $4. I would start at $2 and pay more only when a test on my own work shows a gap.

2. For agents, price the finished task, not the token

   Claude Sonnet 5.5 charges half what Opus 5.5 does per token. Its Terminal-Bench run still cost $7,300 against $4,700 for Opus, because it used 19.4 billion tokens to Opus's 8 billion. GPT-6.1 Sol ran the same 66 tasks for $600 and solved 58.2% of them. For long agent jobs, Sol is the value pick, and Opus earns its premium only on the tasks Sol fails.

3. The cheap end has caught up with April

   GPT-6 Luna costs $0.10 per million input tokens and scores 156.3 on the overall index. Claude Opus 4.7, which Anthropic released in April, scores the same. For tagging, sorting and pulling fields out of records, I would test Luna before paying 20 to 40 times more. It is not a coding agent, though. It solves 16.4% of the Terminal-Bench tasks.

4. What people prefer is not what finishes the work

   The text rating comes from people picking the better of two anonymous answers, between 3,071 and 28,280 votes per model here. That makes it the closest thing to sentiment with numbers behind it, and it has split from capability. Muse Spark 1.3 rates 1494, level with Gemini 3.8 Flash, yet solves 14.5% of the coding tasks. Claude Sonnet 5.5 rates 1471, twelfth of the 17 rated models, while its overall index is fourth. Read the rating as how an answer feels, and check capability separately.

What a finished coding task costs

The price per token hides how many tokens a model uses. Terminal-Bench 4.0 publishes what one full run of its 66 coding tasks cost each model, so dividing that by the tasks the model solved gives a cost per finished task. The cost of failed attempts is already inside that figure. The leaderboard rounds run costs to hundreds of dollars, so treat the result as approximate.

A low cost per task only helps on tasks the model can finish. A model that solves a small share of the set is cheap per success and still fails most of the work. No run cost is published for Muse Spark 1.3.

What the overall index means

The overall index is the Epoch Capabilities Index, which combines a model's results on dozens of benchmarks into one score. Epoch fixes the scale with two reference models. Claude 3.5 Sonnet, from June 2024, scores 130, and GPT-5, from August 2025, scores 150. So 20 points separates two leading models released 14 months apart. Today's top model, Claude Opus 5.5, scores 167.3, which is 17.3 points above GPT-5.

Only the gap between two scores means anything. Each score carries a margin of error of about 3 to 4 points, so two models closer than that are tied.

What changed since September 1, 2026

Each model released since September 1, 2026, compared with the model it replaced. A model's overall index has a margin of error of about 3 to 4 points, so a smaller change is not a real difference.

The large number is the change. A muted change is smaller than the new model's margin of error, so treat it as no real difference.

Full comparison

Select a column heading to sort. The highest score in each column, and the lowest price, is in bold. A hyphen (-) means the evaluator has not tested that model yet, and "n/a" means the maker has not published that figure. Prices are US dollars per million tokens, linked to the page they were read from on October 7, 2026.

Too new to score

These models have no independent scores yet. They will move into the table once an evaluator publishes results.

• Claude Haiku 5.5 (Anthropic, API): $0.10 in / $0.50 out.
• Gemini 3.8 Flash Cyber (Google, limited preview): price not published.
• Mistral Large 4 (Mistral AI, API): price not published.

How to read the numbers

Overall index (Epoch Capabilities Index)

One score that combines a model's results across dozens of benchmarks, run or collected by the research group Epoch AI. The scale is pinned to two reference models, Claude 3.5 Sonnet at 130 and GPT-5 at 150, as explained above. Compare gaps rather than single numbers, and treat a gap inside the margin of error (the range under each score) as a tie.

Agentic coding (Terminal-Bench 4.0)

The share of 66 real command-line tasks an AI agent completes on its own. The benchmark is run by Harbor and the Laude Institute. Most models ran inside their maker's own agent software (Claude Code for Claude, Codex for GPT, Grok Build for Grok). The GLM models also ran in Claude Code, and Gemini ran in a generic agent, which likely understates it. So each score measures the model and the software around it together.

Math (FrontierMath, tiers 1 to 3)

Research-level math problems with checkable answers, kept private and run by Epoch AI. The hardest tier is left out because the top model already solves all of it.

Facts, no search (SimpleQA Verified)

Short factual questions answered with no web access, run by Epoch AI. A wrong answer here is an invented one, which makes it the closest public measure of how often a model makes things up.

Text and web-dev ratings (LMArena)

People compare two anonymous answers and vote, and the votes become a rating. The text rating corrects for length and formatting, so longer answers do not win by default. A gap inside the margin of error (the ± under each rating) is a tie.

Tests we left out, and why

Many widely quoted benchmarks no longer tell the top models apart. A benchmark earns a place here only if it passes three checks. It has results for this season's models, it is run by someone other than the model makers, and the leaders do not all cluster near a perfect score.

• MMLU and MMLU-Pro. Multiple-choice exams across dozens of school and college subjects. In Epoch AI's data the best models already scored near 88% by late 2024, when results stopped being added, and the test no longer tells top models apart.
• GPQA Diamond. Graduate-level science questions written to resist web search (GPQA stands for Graduate-Level Google-Proof Q&A). Every model Epoch AI tested that was released since August scores between 90% and 96%, so a difference here is mostly noise.
• SWE-bench Verified. A test of fixing real bugs in open-source Python projects. OpenAI has published why it no longer evaluates the test, citing contamination, which means models had seen the answers during training. Epoch AI has not run it on any model released after June.
• Terminal-Bench 2. An older version of the coding test used above, and the version Epoch AI's archive still carries. Version 4.0 trimmed the task set to 66 and revised 20 of them, so only 4.0 is cited here.
• AIME. The American Invitational Mathematics Examination, a high-school math competition. Thirty questions a year make one or two lucky answers look like a real gap.
• Humanity's Last Exam. Very hard questions written by subject experts. Epoch AI's imported results cover only three models released since June, too few to compare the field.
• ARC-AGI-2. Puzzles from the Abstraction and Reasoning Corpus that test whether a model can work out a new rule from a few examples. The top four results in Epoch AI's data sit between 93% and 95%, so ranking by it is mostly ranking noise.
• OSWorld. A test of an AI operating a desktop computer. Epoch AI's data has no 2026 results for the original, and none for any model released in September on its successor, OSWorld 2.
• BrowseComp. A test of finding hard-to-locate facts across the web, created by OpenAI. This month's results come from the model makers themselves, and I found no independent run.
• LMArena search. People's votes on answers that use web search. No model released after August has a rating there.

Method and sources

• Benchmark scores come from Epoch AI, Capabilities and benchmarking, and from the LMArena leaderboard dataset published October 2, 2026. Both are shared under the Creative Commons Attribution 4.0 license, which allows reuse with credit. I pulled them on October 7, 2026, kept only the models listed here, kept each model's best result across its reasoning settings, and rounded the values. Model access and license labels, such as "non-commercial", also come from Epoch AI's data.
• Agentic coding results come from the Terminal-Bench 4.0 leaderboard, published on Snorkel AI's site and read on October 7, 2026. The page cites individual results rather than reproducing the leaderboard.
• Prices come from each model's own API pricing page, linked from its input price. Gemini 4 Argon's price is the introductory rate from Google's launch post.
• Where a model has several reasoning settings, the page shows its best result and the price per token, which does not change with the setting. Higher settings use more tokens, so the real cost of a task rises with them.
• Every model release between refreshes is covered in the Pulse, in its own Model and AI releases section each day.