The AI models report. Which model for which task, at what cost.
A weekly, plain-English read on AI models for operations teams: what's best for each kind of work, what's good enough, and what it costs. Tell us your industry and workflow and the report opens with a recommendation for you.
- Free, updated weekly
- The quadrant and the comparison, no signup
- A recommendation for your workflow, with the PDF
Which kind of tool for which kind of task. Two questions decide it.
How much do the inputs vary, and how costly is a mistake? Answer those and the tool picks itself. Most of the money wasted on AI is a frontier model doing a rule's job, or a rule doing a judgment job.
Deterministic automation
Structured inputs, reversible outcomes. A rule on the systems you already own. Cheapest, fastest, easiest to audit; no model needed.
Rules, plus validation and sign-off
Structured inputs, expensive mistakes. The same rule, with a validation layer and a person who signs before anything posts.
Mid-tier model, light review
Open-ended inputs, low stakes. A $$ model drafts; a person sends. Track the edit rate and promote the tool when it earns it.
Frontier model, behind a threshold
Open-ended inputs, high stakes. A $$$ reasoning model with a confidence threshold and a named owner. Starts read-only.
The four quadrants map to our maturity levels: the bottom row is L2 work, the top-left is L2 with governance, the right column is L3. Definitions in the glossary.
Model by task this week.
The best and best-value model for each kind of operations work, and whether a small or open-weight model is enough.
| Task | Best overall | Best value | Small or open-weight enough? |
|---|---|---|---|
| Document extraction Invoices, forms, contracts, PDFs into structured fields | Gemini 2.5 Pro or GPT-5 for layout-heavy pages | Gemini 2.5 Flash / GPT-5 mini | Open-weight vision models for simple, consistent layouts |
| Classification and triage Routing tickets, emails, claims, exceptions | Claude Sonnet 4.5 / GPT-5 | Claude Haiku 4.5 / GPT-5 nano / Gemini Flash-Lite | Fine-tuned small model once you have 2,000+ labelled examples |
| Drafting First drafts of emails, responses, summaries, reports | Claude Sonnet 4.5 / GPT-5 | GPT-5 mini / Claude Haiku 4.5 | Llama 4 or Mistral Medium on-prem where data can't leave |
| Long-document analysis Reading contracts, policies, case files, long threads | Gemini 2.5 Pro (1M context) / Claude Sonnet 4.5 | Gemini 2.5 Flash | Rarely: context length is what you're paying for |
| Reasoning-heavy decisions Multi-step judgment: eligibility, reportability, pricing exceptions | Claude Opus 4.1 / GPT-5 (thinking) / Gemini 2.5 Pro | Claude Sonnet 4.5 / GPT-5 mini (thinking) | No: this is where the frontier matters |
| Coding and automation scripts Integrations, data transforms, glue code, tests | Claude Sonnet 4.5 / Opus 4.1 / GPT-5 | GPT-5 mini / Gemini 2.5 Flash | Open-weight coding models for internal tooling |
| Customer conversations Chat and email with customers under policy | Claude Sonnet 4.5 / GPT-5 | GPT-5 mini / Claude Haiku 4.5 | Only with strict retrieval over your own content |
| Multilingual operations Teams and customers across languages | Gemini 2.5 Pro / GPT-5 | Gemini 2.5 Flash | Open-weight models vary widely by language: test yours |
| On-premises / data can't leave Regulated data, sovereign requirements | Llama 4 Maverick / Mistral Large / DeepSeek V3 (self-hosted) | Llama 4 Scout / Mistral Small / Qwen 3 | Yes: this whole row is open-weight |
| Search and retrieval Finding the right document, precedent or answer internally | Embeddings from OpenAI / Voyage / Cohere + a reranker | Open-weight embeddings (e.g., BGE, E5) self-hosted | Yes |
Names and rankings as of September 15, 2026. Vendors rename and reprice monthly; each edition re-checks. Cost bands rather than prices, on purpose.
The models, side by side.
Every model we currently point clients at, with the band it lands in and what it's actually good for. Context sizes are approximate and change often.
| Provider | Model | Cost band | Context (approx.) | Deployment | Strengths | Best for |
|---|---|---|---|---|---|---|
| Anthropic | Claude Opus 4.1 | $$$ | 200K | API | Deepest reasoning, long agentic tasks, careful with ambiguity | High-stakes decisions behind a threshold; complex multi-step agents |
| Anthropic | Claude Sonnet 4.5 | $$ | 200K (1M beta) | API | Best balance of judgment, speed and cost; strong coding and drafting | Default for drafting, analysis, coding and most agentic work |
| Anthropic | Claude Haiku 4.5 | $ | 200K | API | Fast and cheap with near-Sonnet quality on routine work | Triage, classification, first drafts at volume |
| OpenAI | GPT-5 (thinking) | $$$ | 400K | API | Strong reasoning with adjustable effort; broad tool use | Reasoning-heavy decisions; complex extraction with validation |
| OpenAI | GPT-5 mini | $$ | 400K | API | Good general model at mid cost | Drafting, summarization, routing, coding assistance |
| OpenAI | GPT-5 nano | $ | 400K | API | Very cheap, fast | Classification and simple extraction at scale |
| Gemini 2.5 Pro | $$$ | 1M | API | Largest practical context; strong on documents, video and multilingual | Long-document analysis, layout-heavy extraction, multilingual | |
| Gemini 2.5 Flash | $$ | 1M | API | Fast, cheap for its context size | Document extraction, long-context summaries at volume | |
| Gemini 2.5 Flash-Lite | $ | 1M | API | Cheapest way to read a lot of text | Bulk classification and simple summaries | |
| Meta | Llama 4 Maverick | $$ | 1M | Open-weight | Capable open model; multimodal | On-premises drafting and analysis where data can't leave |
| Meta | Llama 4 Scout | $ | 10M (claimed) | Open-weight | Small, very long context | On-premises triage and retrieval over large archives |
| Mistral | Mistral Large | $$ | 128K | API / self-host | Solid European option with self-hosting | Regulated environments in the EU; drafting and analysis |
| Mistral | Mistral Small | $ | 128K | Open-weight | Cheap, fine-tunable | Fine-tuned classifiers and routers |
| DeepSeek | DeepSeek V3 / R1 | $ | 128K | Open-weight | Strong reasoning at very low cost; self-hostable | Self-hosted reasoning where sovereignty matters; test governance first |
| xAI | Grok 4 | $$$ | 256K | API | Competitive reasoning; real-time data via X | Research and analysis; less common in operations stacks |
| Cohere | Command A | $$ | 256K | API / self-host | Enterprise RAG focus, strong embeddings and rerankers | Internal search and retrieval-grounded answers |
| Alibaba | Qwen 3 | $ | 128K | Open-weight | Strong open model, wide size range, multilingual | Self-hosted multilingual work; fine-tuning |
| Amazon | Nova Pro / Lite | $$ / $ | 300K | API (Bedrock) | Cheap inside AWS; tight IAM integration | Teams already standardized on AWS |
As of September 15, 2026. Open-weight models trade API cost for the people who run them; budget for that.
What it costs, in three bands.
$
Small and fast models (Haiku, Flash-Lite, nano, small open-weight). Pennies per thousand documents. Right for triage, extraction with validation, first drafts.
$$
Mid-tier (Sonnet, GPT-5 mini, Gemini Flash / Pro). The default for drafting, analysis and most agentic work. Ten to fifty times the cost of $ per token.
$$$
Frontier reasoning (Opus, GPT-5 thinking, Gemini Pro thinking). For decisions that need judgment. Use sparingly and behind a threshold; route the easy 80% to cheaper models.
The rule that saves the most money: route the easy 80% of any task to a $ model behind a validation rule. Send only what fails to a $$ or $$$ model. Cost drops by an order of magnitude and quality goes up, because the expensive model only sees the hard cases.