OpenAI vs. Anthropic vs. DeepSeek: A Developer Pricing Breakdown
The three ecosystems price for different strategies. OpenAI spans the widest range: GPT 6 astra sits at $10.00 per million input tokens and $50.00 per million output, while the mini and nano tiers serve high-volume workloads for cents. Anthropic prices Claude as a premium line: Opus at the top for hardest reasoning tasks, Sonnet in the middle as the workhorse, and Haiku competing on speed. DeepSeek competes almost purely on price: DeepSeek Flash costs $0.3 per million input tokens, an order of magnitude below Western flagships, with quality that has closed much of the gap for everyday tasks.
| Strategy | Best fit | Watch out for |
|---|---|---|
| OpenAI | Broad tooling, huge ecosystem, cheap small models | Costs climb fast if you default everything to the flagship |
| Anthropic | Long agentic tasks, coding, strong caching discounts | Opus is premium priced; route only hard tasks to it |
| DeepSeek | Cost-sensitive high volume, batch processing | Rate limits and infrastructure maturity vs. the majors |
List price does not decide it. The shape of your work does
Comparing providers on headline price answers a question nobody has. Every model charges two rates, one for what you send and one for what it writes, and across the 40 models we track output costs a median of 5 times input. So the ranking changes with the job. On our own figures, recorded September 21, 2026, the cheapest model for input-heavy work is GPT 5 Nano, the cheapest for output-heavy work is Ministral 3 3B 2512, and the cheapest for a balanced load is Ministral 3 3B 2512.
That is not a rounding difference in the ordering. Gemini 2.5 Pro ranks 26th cheapest on one workload shape and 31st on another, purely because of how the two rates trade off. Anyone who picked it from a table sorted by input price got a different answer than they would have from the same table sorted for the job they actually run.
The practical consequence: work out roughly how many tokens go in and how many come back for one typical request before you compare anything. That single ratio decides more than the provider does.
Caching is where the providers genuinely differ
Discounted rates for repeated prompt prefixes are the largest published discount on the board, and they are not offered uniformly. 32 of the 40 models we price publish a cache-read rate and 8 do not, so on those 8 there is no such saving to plan for at any volume. Where it exists the median discount is 90 percent off the standard input rate.
The house styles differ too. Anthropic's cache reads run at a median of 10% of its standard input rate across its four models, OpenAI's at 10%, and DeepSeek's between 2% and 3%. If your workload resends a large fixed instruction block on every call, that difference can matter more to the bill than the difference in list price, and it is invisible on a pricing page comparison.
What a price table will not tell you
Three things decide real bills and appear in no comparison. Context limits: a workload whose input does not fit a model's window cannot run on it at any price, which bites hardest on long documents. Image input: a scanned page is not text and every provider tokenises it by its own rule, so the same document costs several times more on one provider than another. Rate limits and maturity: a cheap model you cannot get enough throughput from is not cheap, it is unavailable.
And one more, which is why this site publishes no ranking of quality: we do not benchmark models, we do not scrape anyone's benchmark, and we do not tell you which model is better. We price them. Whether a cheaper model does your job is a question only your own evaluation answers.
Nobody wins on one provider
The winning pattern in production is rarely a single provider. Teams route by task: a cheap model triages and handles the easy majority of traffic, a mid-tier model covers the rest, and a flagship is invoked only when a quality gate fails. Whether that is worth building depends entirely on the spread for your shape of work, and the spread is large. On the support-bot workload priced in our chatbot guide, the same bot costs $6.00 a month on Ministral 3 3B 2512 and $1,080 on GPT 6 astra, a factor of 180x for identical traffic.
Run your own numbers in the cost calculator or compare every model on your shape on the rates page. The gap between the cheapest and the most expensive way to serve the same workload is almost always larger than the gap between the providers.