The last two days of July brought two pricing developments that directly affect the API bill of any organisation using AI. Not roadmap announcements, but hard numbers.
DeepSeek V4-Flash: same price, drastically better
On 31 July 2026, DeepSeek made the official generally available version of V4-Flash-0731 available. What stands out: the model has exactly the same architecture and parameter count (284 billion) as the preview, but DeepSeek only re-ran the post-training, focused on agentic reasoning.
The result is striking. On nine published agent benchmarks, V4-Flash now beats its own flagship V4-Pro-Preview:
| Benchmark | V4-Flash 0731 | V4-Pro-Preview | Flash Preview |
|---|---|---|---|
| Terminal Bench 2.1 | 82,7 | 72,1 | 61,8 |
| DeepSWE | 54,4 | — | 7,3 |
| DSBench-FullStack | 68,7 | — | 37,0 |
The price remained unchanged: $0.14 per million input tokens (cache miss), $0.0028 on a cache hit, $0.28 output. By comparison, on Terminal Bench 2.1, Claude Opus 4.8 scores 85.0, a difference of 2.3 points, at a much higher price point.
For professionals building agentic workflows (automated document processing, code reviews, data extraction), this is a significant shift: you now get Pro-level quality at Flash rates, without changing your API calls.
OpenAI: GPT-5.6 Luna 80% cheaper
A day earlier, on 30 July, OpenAI announced via CNBC that GPT-5.6 Luna is becoming 80% cheaper. The new rates:
- Luna input: $0.20/M tokens (was $1.00)
- Luna output: $1.20/M tokens (was $6.00)
- Terra input: $2.00/M tokens (was $2.50, -20%)
- Sol: unchanged
OpenAI attributes the reduction to efficiency gains in the inference process, partly achieved by having Sol itself rewrite its own inference stack. Luna and Terra can now handle workloads that previously required a more expensive model.
What does this mean for your choice of tools? Luna is now cheaper than Claude Sonnet 4.6 ($3.00/M input) and cheaper than DeepSeek V4-Flash for output, but more expensive for input. Terra enters the segment of GPT-class models for structured processing tasks.
What do you change now?
A few practical considerations:
Agentic pipelines: DeepSeek V4-Flash is now the most cost-efficient choice as an agent backbone, provided you accept the GDPR considerations of a Chinese model (processing outside the EU, no published SCCs). Organisations with strict data location requirements would do better to stay with EU-hosted models.
Batch processing and extraction: OpenAI Luna at $0.20/M input is now attractive for large document volumes where you previously found GPT-5.6 Sol too expensive. Check whether accuracy meets your use case before you switch.
Cost calculation when switching: weigh the benchmark differences against your actual workload. Terminal Bench 2.1 measures agentic coding tasks; for legal summarisation or customer service replies, the ratios may differ.
For a full overview of prices per model, see our models page. You'll find sector-specific professional tools under /voor/.
Sources: DeepSeek blog – V4-Flash GA · CNBC – OpenAI price reduction · American Bazaar – Luna -80% · Cosmic JS – V4-Flash benchmarks