The three are not competing for the same job
Anthropic and OpenAI both sell general-purpose frontier intelligence through an API. Cohere sells something narrower: models tuned for enterprise retrieval, classification and multilingual search, with deployment options the other two do not offer. Comparing them on a single benchmark misses what actually separates them.
Current pricing
| Model | Provider | Input /1M | Output /1M | Context |
|---|
| Claude Opus 5 | Anthropic | $5 | $25 | 1M |
| Claude Sonnet 5 | Anthropic | $2 (intro) | $10 (intro) | 1M |
| Claude Haiku 4.5 | Anthropic | ~$0.80 | ~$4 | 200K |
| GPT-5.6 Sol | OpenAI | $5 | $30 | ~1.05M |
| GPT-5.6 Terra | OpenAI | $2.50 | $15 | ~1.05M |
| GPT-5.6 Luna | OpenAI | $1 | $6 | ~1.05M |
| Command R+ | Cohere | Enterprise pricing | Enterprise pricing | 128K |
Claude Sonnet 5 introductory pricing runs to 31 August 2026, then moves to $3/$15. Cohere prices Command R+ through enterprise agreements rather than a public per-token rate — check current terms directly.
✓ Quick answer
- Best for agentic coding: OpenAI — Codex delivers a pull request from a task description
- Best for writing and long output: Anthropic — Opus 5 at 128K output, best prose quality
- Best for enterprise RAG and search: Cohere — Rerank and Embed are purpose-built for this
- Best for private deployment: Cohere — VPC and on-prem options the other two do not match
- Best safety posture for procurement: Anthropic — highest FLI grade at C+
- Broadest modality coverage: OpenAI — text, image, video, voice under one account
Anthropic
Three tiers, all with a 1M context window on the current line: Opus 5 for hard reasoning and long-form output, Sonnet 5 as the everyday workhorse, Haiku 4.5 for classification and routing at speed. Opus 5 carries an effort dial that lets you trade cost against depth per request, which neither competitor exposes.
Practical strengths: the best writing quality of the three, a 128K output window on Opus 5 that matters for document generation, prompt caching that cuts repeated-context cost substantially, and a Batch API at half price for work that does not need real-time responses. Anthropic also holds the highest published safety grade (FLI C+), which matters in regulated procurement more than most engineering teams expect.
Weakness: no native image or video generation. If your application needs those, you are integrating a second provider regardless.
OpenAI
The broadest platform of the three. GPT-5.6 in three tiers, plus Codex for agentic coding, DALL-E for images, Sora for video, the Realtime API for voice, and Whisper for transcription — all under one account and one billing relationship.
Codex is the differentiator that has no equivalent elsewhere: describe a coding task, and it works in a persistent cloud session and returns a pull request. For teams with a backlog of well-defined tickets, that workflow is worth more than a few benchmark points. GPT-5.6 Sol also leads Terminal-Bench 2.1 at 88.8%.
Watch the tier gap. Luna at $1/M is excellent for routing and classification but drops sharply on multi-step reasoning — 41.3% on Nerova against Terra's 71.4%. Do not use Luna for chained reasoning simply because it is cheap.
Cohere
Cohere is the odd one out, and deliberately so. Command R+ is built for retrieval-augmented generation rather than open-ended chat: it is tuned to cite sources accurately, follow structured output formats, and work well inside a RAG pipeline rather than replace one.
The wider Cohere product set is where the real differentiation sits. Embed and Rerank are strong at multilingual semantic search, and Rerank in particular is used by teams who run a different model for generation but want Cohere handling the retrieval quality. Cohere also offers private cloud and on-premise deployment options that Anthropic and OpenAI do not match at the same level, which is the deciding factor for organisations with strict data-residency requirements.
It is not the choice for general-purpose intelligence. If you want the smartest model, it is not Cohere. If you want the best enterprise search and retrieval layer with deployment flexibility, it frequently is.
Which to build on
| If you are building… | Use | Why |
|---|
| An agentic coding tool | OpenAI | Codex async PR delivery; Terminal-Bench 88.8% |
| Document generation or analysis | Anthropic | 128K output, 1M context, best prose quality |
| Enterprise search over private data | Cohere | Embed + Rerank; multilingual retrieval quality |
| Multimodal (text, image, voice, video) | OpenAI | Only provider covering all four natively |
| High-volume classification or routing | Anthropic Haiku or GPT-5.6 Luna | Both optimised for latency at low cost |
| A deployment inside your own VPC | Cohere | Private and on-prem options the others lack |
| Anything in a regulated industry | Anthropic | FLI C+, published safety framework, BAA available |
The practical answer for most teams
Most production applications end up on two providers, not one — typically Anthropic for text-heavy reasoning and OpenAI for coding and multimodal, with Cohere added if retrieval quality becomes the bottleneck. The SDKs are close enough that switching a model string is usually a one-line change, so building with a provider abstraction from the start costs little and preserves the option.
Frequently asked questions
Which LLM API is cheapest?
Of the three, GPT-5.6 Luna at $1/$6 per million tokens is the lowest published rate, followed by Claude Haiku 4.5 at roughly $0.80/$4. Cohere prices through enterprise agreements rather than a public rate card. Note that cheaper tiers degrade on multi-step reasoning — Luna scores 41.3% on Nerova against Terra's 71.4%.
Can I switch between these APIs easily?
Mostly yes for basic completion calls — the request shapes are similar enough that a provider abstraction layer makes switching a configuration change. Where it gets harder is provider-specific features: Codex, prompt caching, Cohere Rerank and the effort dial on Opus 5 have no cross-provider equivalent.
Which is best for RAG applications?
Cohere for the retrieval layer — Embed and Rerank are purpose-built for it and are commonly used alongside a different generation model. For generation over retrieved context, Claude Sonnet 5's 1M context window means many applications can skip chunking entirely and pass whole documents.