AI coding costs are falling fast in 2026. Price cuts from the big labs, discounts for large customers and strong open-weight models that companies can run themselves have all pushed the cost of AI-assisted programming down. This guide pulls together the latest numbers and real cases from AI Newsround reporting so teams can compare options, with a checklist at the end.
Key takeaways
- OpenAI's GPT-6.1 Terra is priced at $1 per million input tokens and $5 per million output tokens.
- Mistral's Large 4 open weights have an API price of $0.70 and $2.10 per million tokens, with a licence catch for resellers.
- Revolut says it cut coding inference costs about 90% by self-hosting an open-weight model for its engineers.
- Reports of steep discounts between OpenAI and Anthropic are real but unconfirmed, and "up to 90%" is one customer's saving, not a price list.
Option 1: Frontier API models
OpenAI says GPT-6.1 Terra matches its flagship on one benchmark at one-tenth the price per token, though that is the vendor's own test. Early independent testing found a smaller, older model was cheaper and faster on very simple tasks. Read the details in our GPT-6.1 Terra explainer.
Option 2: Negotiated enterprise discounts
Large customers such as Rippling and Replit reportedly moved spending after OpenAI offered aggressive terms, according to The Information. See what is verified about the discounts. Smaller teams rarely get these terms, but the competition pushes list prices down.
Option 3: Open-weight models
Mistral released Large 4, a 1-trillion-parameter model, under a modified MIT licence. Businesses selling hosted access need a separate licence, and the model needs serious GPU hardware. See our Mistral Large 4 guide.
Option 4: Self-hosting at scale
Revolut moved about 2,000 engineers to a self-hosted Chinese open-weight model running on 56 Nvidia B200 chips, reporting about $0.30 per million tokens. The figure covers inference costs, and the coverage does not say how upfront hardware costs are counted. Read the Revolut case.
Hidden costs to budget for
- Tool risk: Meta bought Conductor and is closing its cloud service on 20 November; see the details.
- Security: developers reported a desktop app silently installing agent plugins; see what to check.
- Governance: Goldman Sachs limited coding agents to engineers; see the report.
Checklist: how to pick an AI coding setup
- Measure cost per completed task on your own code, not tokens alone.
- Test at least one commercial and one open-weight model on the same tasks.
- Route simple jobs to cheaper models and hard jobs to stronger ones.
- Read licences, especially for hosted or resale use.
- Audit agent settings and plugins, and keep copies of your configuration.
- Re-test every few months, as prices and quality keep changing.
Frequently asked questions
What is the cheapest way to use AI for coding?
It depends on volume. Light users usually save most with a cheaper API model; heavy users may gain from self-hosting, as Revolut did, but must count hardware and staff time.
Are open-weight models good enough for coding?
Revolut reports strong results from one, with routing for image work. Test on your own tasks rather than relying on benchmark claims.
Is the price war permanent?
Nobody knows. Financing pressure on AI companies could change pricing; see our analysis of the AI sell-off.
Anthropic, the maker of Claude Code, is the company behind this assistant. This guide summarises published reporting and does not recommend a vendor.
Sources: OpenAI, Mistral AI, The Register, The Information via The Decoder, TechCrunch, Cybernews, as cited in the linked AI Newsround reports.
