Best LLMs in 2026: the right model for each task

There is no single best LLM in 2026. The best large language model for a coding agent, for reviewing a 600,000-token contract set and for labelling ten million support tickets a day are three different models, and the reason is usually one published fact: a context window, a prompt-length surcharge, a license or a rate that halves at night. This guide gives the strongest candidates for each kind of work, the verified facts behind each pick, and a method for deciding which LLM is best on your own workload.

Every rate below is the vendor's own list rate, read on its pricing page on 19 October 2026; a negotiated rate on your own account replaces it. For the full tables by modality, see the AI models comparison. This page is the recommendation; that one is the data.

Why "which LLM is best" has no single answer

Benchmarks average a model's performance over tasks that are not yours. Rates differ by a factor of a hundred between tiers. And the facts that rule a model in or out are specific: whether 600K tokens fit, whether the license allows your use, whether the model is a preview. So the useful question is not "which model is best" but "which model is best for this task, at this volume, under these constraints".

The picks below come from three kinds of evidence, in this order: what the vendor says the model is for, the published facts that make it fit or not, and only then public leaderboards, quoted with their date. None of them replaces running the candidates on your own inputs.

The best LLMs by task in 2026

Best LLMs for coding

Anthropic positions Claude Opus 5.5 for long-running agentic coding and knowledge work, and recommends starting most workloads with it¹. It lists $4 input and $20 output per million tokens, with cache hits at $0.20, a 1M window and 128K output²¹. For a lower rate, Moonshot's Kimi K2.7 Code lists $0.95 input, $0.19 cached and $4 output with a 262,144-token window³; MiniMax M3 lists $0.30 and $1.20 up to 512K input tokens⁴; xAI's Grok Build 0.1 lists $1 and $2 with a 256K window⁵.

Why caching decides: a coding agent resends the repository map, tool definitions and open files on every step, so after the first call most input tokens are cache reads. Compare the cached rate and the context limit before the headline rate. The coding models comparison works through prompt-length tiers on real step sizes, and the decisive test is your own: tasks from your repository, judged by the test suite.

Best LLMs for long context

Here the deciding fact is how each vendor bills a large prompt, not the window size alone. Anthropic bills Claude 4.6 and later models, except Haiku 5.5, across the full 1M window at the standard rate². OpenAI reprices a GPT-6 request above 272K input tokens: Astra goes to $20 input and $75 output for that request⁶. Google lists Gemini 3.1 Pro at $4 and $18 above 200K prompt tokens, with a 1,048,576-token input limit⁷⁸. Kimi K3 has a 1,048,576-token window at $3 and $15³. Grok 4.7's window is 500K⁵, so a 600K prompt does not fit.

A large window is not a reason to fill it. The guide to managing long context windows covers context rot, long-context price tiers and a token budget with compaction rules.

Best LLMs for the hardest reasoning

Each vendor names its pick. Anthropic describes Claude Fable 5.1 as being for demanding reasoning and long-horizon agentic work, and suggests it when evals on Opus 5.5 at higher effort still fall short¹. OpenAI calls GPT-6 Astra its most capable model for the most demanding work, and GPT-6.1 Sol near-Astra performance at a lower cost⁹. Fable 5.1 and Astra list the same $10 input and $50 output; GPT-6.1 Sol lists $2 and $10²⁶.

The honest test is narrow: take the requests your current model gets wrong, run the candidates on those, and see which fixes enough of them to justify its rate. The Fable 5.1 vs GPT-6 Astra comparison sets the two side by side.

Best LLMs for cheap, high-volume work

Classification, extraction, routing and short summaries are where the bill concentrates. Two models start at $0.10 per million input tokens and $0.50 output: OpenAI's GPT-6 Luna, which OpenAI calls its most efficient model for focused, high-volume tasks⁹⁶, and Anthropic's Claude Haiku 5.5, described for classification, extraction and routing¹. Their conditions differ: Haiku 5.5 bills a prompt over 100,000 tokens at $0.50 and $2.50 for the whole request², while Luna's higher rate starts above 272K⁶.

Two more have conditions worth calendaring. Gemini 3.8 Flash lists $0.75 and $3.75 through 31 December 2026 and twice that from 1 January 2027⁷, with a 1,048,576-token input limit and audio, video and PDF input¹⁰. DeepSeek Flash lists $0.30 and $1.20 at peak and half that off-peak, outside 01:00–04:00 and 06:00–10:00 UTC on weekdays¹¹. GLM-5.3 Flash ($0.15 / $0.50) and Qwen3.8 Flash ($0.15 / $0.47) sit in the same tier¹²¹³.

At this volume, a few points of pass rate move cost per successful task more than a few cents of rate, because every failure is retried somewhere more expensive. The small and fast models comparison has the arithmetic.

Best open-weight LLMs for self-hosting

If the requirement is that prompts never leave your infrastructure, the model must publish weights under a license that permits your use. OpenAI's gpt-oss-120b is Apache 2.0, has a 131,072-token context and, by OpenAI's description, fits on one H100 GPU¹⁴. Mistral AI lists Mistral Large 3 as an open-weight, general-purpose multimodal model and Mistral Small 4 as a hybrid instruct, reasoning and coding model, both Apache 2.0¹⁵. The Qwen team publishes Qwen3.8 models on Hugging Face, from a 27B model to a 2.4T-parameter mixture-of-experts, each with an FP8 variant¹⁶; the organization page does not state their license, so read each model card.

Self-hosting moves the work from the rate card to operations: serving stack, quantization, throughput and the tool-call parser. The guides on self-hosting with vLLM and Chinese open-weight models in production cover both.

Best LLMs for multilingual work

Vendors state multilingual support broadly. Anthropic says all current Claude models support multilingual capabilities¹; Cohere describes Aya Expanse as a multilingual model that works with 23 languages¹⁷. Neither statement tells you how a model handles your languages and your domain, so this is the task where a small evaluation set per language matters most.

One cost detail is easy to miss: tokenizers differ, so the same text is a different number of tokens on different models. Anthropic notes that 1M tokens is roughly 555,000 English words on its current tokenizer, against about 750,000 on models before it¹. Count tokens per document in each of your languages before comparing per-token rates.

Best LLMs for agents and tool use

An agent is only as good as its tool calls. Anthropic states that all current Claude models support tool use and positions Opus 5.5 for long-running agentic work¹. Google publishes a second Gemini 3.1 Pro model code, gemini-3.1-pro-preview-customtools, beside the standard one⁸. Mistral describes Mistral Medium 3.5 as optimized for agentic and coding use cases¹⁵.

What decides is reliability on your schemas: does the model call the right tool, with valid arguments, and stop when it should. Test that with your real tool definitions; the guides to tool calling in production (argument validation, timeouts, result caps) and bounding agent loops (step, deadline and spend limits) cover the controls around it.

What public leaderboards say, and what they don't

Leaderboards are useful for a first shortlist. On the Artificial Analysis leaderboard opened on 19 October 2026, the highest entry on its Intelligence Index was Claude Opus 5.5 at its maximum effort setting, at 58, followed by Claude Sonnet 5.5 at maximum effort at 56; GPT-6 Astra at maximum effort and Claude Fable 5.1 scored 53¹⁸. The page showed no date for the ranking itself.

Two things in that table matter more than the order. First, the same model appears several times: Opus 5.5 scored 58 at maximum effort, 56 at xhigh and 54 at high¹⁸. Effort is a setting you choose, and it changes both the score and the bill. Second, an index is an average across tests that are not your task. The LMArena leaderboard could not be opened for this article and is not quoted.

How to choose the best LLM for your workload

1. Write down the task and its check. A passing test suite, a schema-valid extraction, a label that matches, a reviewer's accept. Without a check there is no "best".
2. Shortlist on hard facts. Window, output limit, input modalities, region, license, preview or retiring status. The tables in the AI models comparison are the place to start.
3. Collect 50 to a few hundred real inputs per task type, including the hard ones and the ones that must be refused. The guide to evaluating LLMs on real workloads covers sampling and labelling.
4. Run each candidate under identical conditions: same prompt, tools, effort and caching. Record estimated provider cost and outcome per request, retries included.
5. Compare cost per successful task, not rate per token. See cost per successful task.
6. Assign a model per task tier, with a tested fallback. A capable model for hard requests and a small one for volume is the common outcome; quality-aware routing covers sending each request to the right tier.
7. Re-run on a trigger. A new model, a rate change such as Gemini 3.8 Flash's on 1 January, or a retirement date is a reason to repeat the evaluation.

Choosing and switching models with INFRO

The best LLM this quarter may not be next quarter's, so the choice should live in configuration. INFRO is a control plane in front of the provider accounts you connect: you call every candidate through one OpenAI-compatible endpoint on your own keys, and your providers bill you directly at your rates. INFRO's subscription is separate and is not priced per token.

• Evaluate. Each response carries usage.cost, an estimate of what your provider will bill, and GET /v1/usage?group_by=model returns cost, error rate and latency percentiles per model for an evaluation run (usage analytics); traces group the steps of an agent run (request traces).
• Decide per task. A route such as code-agent or ticket-labels lists the chosen model first and a tested alternative after it, and fails over to the next target when a provider errors (routing).
• Govern. Model access allowlists limit which models can be called, so a team cannot drift to an unapproved model.

When a better candidate appears, you run it on the same evaluation set and change the route's targets; the application keeps sending the same route name.