Two rules of thumb before the recommendations. First, bigger models are better and dearer: a larger model follows nuanced instructions more reliably and needs fewer retries, but you pay for it on every token — so use the smallest model that does your job well, not the best model on a leaderboard. Second, UK-hosted is a residency choice, not a performance one: our UK-hosted models cost more per token than the cheapest partner routes, and in exchange your prompts never leave our infrastructure. All prices below come from the live catalogue and update daily (last generated 2026-07-08).

Customer support automation

Support traffic is high-volume and mostly routine, so cost per token dominates. Use a small model to classify and answer the routine 80%, and escalate the hard 20% to a bigger model or a human with full context.

  • GPT OSS 20B — $0.04 (£0.03) in / $0.16 (£0.12) out per 1M tokens. The volume workhorse: cheap enough to run on every ticket, with function calling to look up orders and reasoning for the awkward ones.
  • Llama 3.3 70B Instruct — $0.12 (£0.09) in / $0.37 (£0.28) out per 1M tokens. The escalation tier: noticeably better on multi-step customer problems, still a fraction of frontier prices.
  • Qwen3.6 27B UK-hosted — $0.32 (£0.24) in / $0.64 (£0.48) out per 1M tokens. The residency option: when tickets contain personal data you'd rather keep on our own hardware, in one jurisdiction, end to end.

Document drafting and summarising

Drafting rewards instruction-following quality — a better model saves editing time, and drafting volumes are usually low enough that the per-token premium doesn't hurt.

  • Qwen3.6 27B UK-hosted — $0.32 (£0.24) in / $0.64 (£0.48) out per 1M tokens. Our default drafting model: strong general writing, and prompts (your unreleased contracts, board papers, client letters) never leave our UK datacentre.
  • GPT OSS 120B — $0.04 (£0.03) in / $0.19 (£0.14) out per 1M tokens. A step up in reasoning and structure for long or technical documents, at open-weight prices.
  • Qwen3 235B A22B Instruct 2507 — $0.11 (£0.08) in / $0.12 (£0.09) out per 1M tokens. Large mixture-of-experts generalist with a 262K context — good when the 'draft' means digesting a pile of source documents first.

Code assistant

Split the job in two: completion wants a small fast model called thousands of times a day; implementation and review want the biggest code model you can justify, called occasionally.

  • Qwen2.5 Coder 14B UK-hosted — $0.16 (£0.12) in / $0.32 (£0.24) out per 1M tokens. For autocomplete and boilerplate: fast, cheap, and your source code stays on our UK infrastructure.
  • Qwen3 Coder 480B A35B Instruct — $0.35 (£0.26) in / $1.17 (£0.88) out per 1M tokens. For real implementation work: multi-file changes, debugging and review, with function calling for agentic coding tools.
  • Qwen2.5 Coder 32B Instruct — $0.77 (£0.58) in / $1.17 (£0.88) out per 1M tokens. The middle path when the 14B isn't quite enough and the 480B is overkill.

A RAG pipeline needs two models: an embedding model to index and retrieve your documents, and a chat model to compose the answer from what was retrieved. Embedding costs are almost negligible — you pay input tokens only, once per document plus once per query. One honest caveat: our embedding models are currently all partner-routed; the UK-hosted set doesn't include one yet, so document text sent for embedding does leave our infrastructure.

  • BGE M3 — $0.01 (£0.01) in / — per 1M tokens. Our default embedding recommendation: multilingual, strong retrieval quality, effectively free at RAG volumes.
  • Qwen3 Embedding 0.6B — $0.01 (£0.01) in / — per 1M tokens. An equally cheap alternative with strong benchmark scores, if you want to A/B retrieval quality.
  • Llama 3.3 70B Instruct — $0.12 (£0.09) in / $0.37 (£0.28) out per 1M tokens. The answering model: good grounding behaviour at mid-range prices; pair with a UK-hosted model instead if the retrieved content is sensitive.

Transcription

Speech-to-text is priced per minute of audio, and every model in the audio category costs well under a penny a minute — so accuracy on your audio, not price, should decide.

  • Whisper Large v3 — $0.0006 (£0.0004) in / — per minute of audio. The default: excellent multilingual accuracy on meetings, calls and recorded media.
  • Whisper — $0.0005 (£0.0004) in / — per minute of audio. The budget option for clean audio at bulk volumes — older model, lowest per-minute price in the catalogue.
  • Nova 3 (speech-to-text) — $0.01 (£0.008) in / — per minute of audio. Built for real-time conversational audio (voice agents, live calls) rather than batch files.

Image and document understanding

Vision-language models read scans, screenshots, charts and photos. Match the model to the stakes: extraction from clean documents is easy; interpreting messy real-world images is not.

  • Llama 4 Scout 17B 16e Instruct — $0.12 (£0.09) in / $0.35 (£0.26) out per 1M tokens. The value pick for everyday document and image Q&A, with function calling and a huge context window.
  • Qwen3 VL 235B A22B Instruct — $0.23 (£0.18) in / $1.03 (£0.77) out per 1M tokens. The quality pick for hard visual work: dense documents, small print, charts and diagrams.
  • Llama 3.2 11B Vision Instruct — $0.40 (£0.30) in / $0.40 (£0.30) out per 1M tokens. A smaller, steady choice for high-volume image tagging and description.

Agents and function calling

Agents multiply token spend — every tool call adds another round trip through the model — so you want function calling, decent reasoning and a price you can afford in a loop. Look for the Function calling badge in the catalogue.

  • GPT OSS 120B — $0.04 (£0.03) in / $0.19 (£0.14) out per 1M tokens. The best balance we serve for agent loops: strong reasoning, function calling, and cheap enough to iterate.
  • Qwen3 30B A3B — $0.06 (£0.04) in / $0.39 (£0.29) out per 1M tokens. The budget agent: mixture-of-experts efficiency makes multi-step loops very cheap; escalate when it gets stuck.
  • DeepSeek R1 Distill Qwen 32B — $0.58 (£0.44) in / $5.71 (£4.27) out per 1M tokens. For agents that plan: distilled chain-of-thought reasoning — note reasoning models bill more output tokens by design.

When only the best will do (frontier models)

The catalogue also carries the frontier proprietary models — the ones at the top of the leaderboards. Two honest caveats before you reach for them. They cost an order of magnitude more per token than the open-weight models above, and because we aggregate them through partner routes, our price includes routing fees and our margin — if you are running frontier-model volume all day, buying direct from the vendor will be cheaper, and we will tell you so. And like every partner-routed model, prompts to them leave our infrastructure. Where they earn their keep is the hard 5% of your workload: route the bulk to an open-weight model and escalate only the requests that defeat it.

  • Claude Sonnet 4.6 — $3.51 (£2.63) in / $17.55 (£13.13) out per 1M tokens. The dependable all-rounder for hard drafting, analysis and agentic work.
  • GPT 5.1 — $1.54 (£1.15) in / $12.34 (£9.24) out per 1M tokens. Strong general reasoning and tool use across the board.
  • Gemini 3.1 Pro — $2.34 (£1.75) in / $14.04 (£10.51) out per 1M tokens. A 1M-token context window for whole-repository or whole-case-file jobs.

Compliance-sensitive workloads

If your prompts contain special-category data, unreleased financials or anything your regulator or your client contracts say must stay within one jurisdiction, route them to the UK-hosted models — they keep the work in a single jurisdiction on our own hardware: GPUs we own in our UK datacentre, with prompts never leaving our infrastructure. That residency carries a premium over the cheapest partner routes, and we'd rather you pay it knowingly than discover the difference in an audit.

  • Qwen3.6 27B UK-hosted — $0.32 (£0.24) in / $0.64 (£0.48) out per 1M tokens. UK-resident general assistant: drafting, summarising and analysis with no third-party AI provider in the data path.
  • Qwen2.5 Coder 14B UK-hosted — $0.16 (£0.12) in / $0.32 (£0.24) out per 1M tokens. UK-resident code assistant for proprietary source code.

For everything else in this guide, the partner-routed models are processed on a vetted partner provider's infrastructure — fine for most workloads, but it is a data flow you should record in your processing register. Every model page states its residency class explicitly.

Guard rails for user-facing AI

If the public can type into your model, put a safety model in front of it. They're cheap, and they run as a second, parallel call.

  • Llama Guard 4 12B — $0.21 (£0.16) in / $0.21 (£0.16) out per 1M tokens. Screens prompts and responses against a policy taxonomy before they reach your users.
  • Nemotron Content Safety 3.5 — $0.23 (£0.18) in / $0.23 (£0.18) out per 1M tokens. An alternative safety classifier if you want different policy coverage or a second opinion.

Browse the full catalogue of 305 models, see how the gateway works, or sign up and test your shortlist with £15 of free credit — the whole point of per-token pricing is that trying three models costs pennies.

Still not sure?

Tell us what you're building and an engineer will recommend a model, sanity-check your token budget and get you calling the gateway the same day.