DeepSeek V4-Flash-0731: agentic coding at $0.14/M
DeepSeek shipped V4-Flash-0731 on July 31 under an MIT license — 82.7 on Terminal Bench 2.1, $0.14 input and $0.28 output per million tokens. Route your cheap work here.
DeepSeek updated its cheap tier on July 31, and the interesting number isn't the price — the price didn't move. It's the benchmark. DeepSeek V4-Flash-0731 posts agent-and-coding scores in the same neighborhood as models that cost dramatically more per token, and it ships under an MIT license. If you run agents on a budget, your routing table just changed.
What actually happened
The model card reports 82.7 on Terminal Bench 2.1, 54.4 on DeepSWE, 70.3 on Toolathlon-Verified, 68.7 on DSBench-FullStack, and 76.7 on Cybergym. It's a mixture-of-experts model at 304B total parameters with a million-token context window, and DeepSeek recommends a maximum output length of 384K tokens at higher reasoning effort. The license is MIT — commercial use, self-hosting, no asterisk.
Pricing on the DeepSeek API is unchanged: $0.14 per million input tokens on a cache miss, $0.0028 on a cache hit, and $0.28 per million output tokens. Artificial Analysis scores the release at 50 on its Intelligence Index.
Two caveats before you rewire anything. Terminal Bench and DeepSWE measure agentic coding tasks, not your codebase — every model on that leaderboard was tuned with the leaderboard in mind. And the API runs on DeepSeek's infrastructure in China, which is a data-residency decision, not just a latency one.
Why cheap open weights matter for your business
Most agent workloads are not hard. Reading a ticket, classifying it, pulling three fields out of a PDF, writing a migration script, summarizing a support thread — that's the bulk of the token spend at a small company, and it's the part that quietly runs 24/7. Paying frontier-model rates for it is a choice, not a requirement.
The move is a routing layer, not a migration. Keep the expensive model for the work that actually needs judgment. Send the volume — retries, classification, first-pass drafts, test generation — to a cheap tier, and measure the failure rate rather than assuming one. We've argued this before as cost per completed task, and this release makes the arithmetic more lopsided than it was last week.
The MIT license is the part that outlasts the benchmark. Open weights mean the model can't be retired out from under you, repriced at renewal, or pulled by policy. You can pin a version, run it on your own hardware, and keep it running after the vendor moves on. That's worth more to a ten-person company than three points on a benchmark.
Key takeaways
- DeepSeek V4-Flash-0731 shipped July 31, 2026 under an MIT license, with a 1M-token context window
- Reported scores: 82.7 Terminal Bench 2.1, 54.4 DeepSWE, 70.3 Toolathlon-Verified — agentic and coding workloads
- Price held at $0.14 / $0.28 per million tokens, $0.0028 on cache hits
- Benchmarks aren't your workload — test it on your actual tasks and track failure rate, not vibes
- Open weights mean the model can't be retired, repriced, or geoblocked out from under your product
We build the routing layer, not the lock-in. Cheap models for volume work, frontier models for judgment, and a swap that takes an afternoon instead of a quarter. See how we build portable AI systems, or run your token math before you commit to a vendor's price list.
Sources: DeepSeek model card (Hugging Face), DeepSeek API pricing, Artificial Analysis.
- #deepseek
- #open-weights
- #agentic-coding
- #token-costs
- #model-routing
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
ShieldFont poisons AI scrapers. It poisons buyers too.
ShieldFont is a web font that feeds AI scrapers gibberish while humans see real text. Clever — and exactly wrong if you sell through AI search.
Read itCisco Antares: security models small enough to run local
Cisco open-weighted two tiny models that find where a CVE lives in your codebase — Apache 2.0, 350M and 1B params, and your source never leaves the building.
Read it