DeepSeek V4 Flash 0731 is the production release of DeepSeek's small agentic model, shipped on July 31, 2026, that scores 82.7 on Terminal-Bench 2.1 while charging $0.14 per million input tokens and $0.28 per million output tokens. It beats DeepSeek's own 1.6-trillion-parameter Pro model on agent benchmarks, lands within roughly three points of GPT-5.6 Sol, and costs about 1/35th as much on input. If you run a coding agent in production, this is the release that changes your unit economics.
The number that should stop you is not 82.7. It is the delta: the preview build of the same model scored 61.8. Same architecture, same parameter count, same everything — DeepSeek moved a model 21 points on a hard agentic benchmark purely through training and post-training work. That is a very loud statement about where the remaining headroom in small models actually lives.
Below: the full benchmark picture, real pricing math against the frontier labs, what "same architecture" actually means, and the honest case for not switching.
Key Takeaways
- DeepSeek V4 Flash 0731 scores 82.7 on Terminal-Bench 2.1, up from 61.8 for the preview build and 72.1 for V4-Pro-Preview.
- Pricing is $0.14/M input on cache miss, $0.0028/M on cache hit, and $0.28/M output, with a 2,500-request concurrency limit.
- Architecture is unchanged from the preview: 284B total parameters, 13B active, 1M-token context. This was a training upgrade, not a new model.
- It beats Z.AI's GLM-5.2 (81.0) and trails Claude Opus 4.8 (85.0) by about two points on the same benchmark.
- The release added Codex compatibility, making it a near drop-in for existing agent harnesses.
What is DeepSeek V4 Flash 0731?
DeepSeek V4 Flash 0731 is the 0731-dated production checkpoint of the Flash tier in DeepSeek's V4 family, released on July 31, 2026 after a public preview period. The V4 line launched in April 2026 as two models: V4 Pro at 1.6 trillion total parameters with roughly 49 billion active per token, and V4 Flash at 284 billion total with roughly 13 billion active. Both share a one-million-token context window.
DeepSeek has been unusually explicit that this release keeps the architecture and parameter count identical to the preview. As MarkTechPost reported, the gains come from training and post-training, not from a bigger network. For anyone who has been watching labs quietly inflate parameter counts to chase leaderboards, that transparency is worth something on its own.
We covered the underlying architecture when the family launched, in DeepSeek V4 and hybrid sparse attention — the mechanism that makes a 1M-token window affordable at this price tier is the same one doing the work here.
How does DeepSeek V4 Flash compare on benchmarks and price?
On Terminal-Bench 2.1, DeepSeek V4 Flash 0731 scores 82.7, placing it above Z.AI's GLM-5.2 at 81.0 and below Claude Opus 4.8 at 85.0. On price it is not close to anything — at $0.14 input and $0.28 output per million tokens, it undercuts every frontier model by more than an order of magnitude.
| Model | Terminal-Bench 2.1 | Input $/M | Output $/M |
|---|---|---|---|
| Claude Opus 4.8 | 85.0 | — | — |
| GPT-5.6 Sol | ~3 pts above Flash (reported) | $5.00 | $30.00 |
| DeepSeek V4 Flash 0731 | 82.7 | $0.14 | $0.28 |
| GLM-5.2 (Z.AI) | 81.0 | — | — |
| DeepSeek V4 Pro (preview) | 72.1 | — | — |
| DeepSeek V4 Flash (preview) | 61.8 | — | — |
Run the arithmetic on a realistic agent workload. A coding agent that burns 8M input and 2M output tokens per day costs about $100/day on GPT-5.6 Sol at list price. The same workload on V4 Flash 0731 costs about $1.68. Over a month that is roughly $3,000 versus $50 — and with cache hits at $0.0028/M, a workload with a large stable system prefix goes lower still.
For context on what you give up at the top of the market, see our breakdowns of Claude Opus 5 and Grok 4.5. The frontier is still the frontier. The question in 2026 is how often you actually need it.
Why did a training-only update gain 21 points?
Because Terminal-Bench measures agentic competence, not knowledge, and agentic competence is mostly a post-training problem. The benchmark runs a model inside a real terminal against real multi-step tasks — reading output, recovering from errors, deciding when a task is done. A model can know everything about bash and still fail because it does not know when to stop retrying.
That skill set responds enormously to reinforcement learning on trajectories, tool-use formatting, and failure-recovery data. None of it needs more parameters. The 61.8 to 82.7 jump is the clearest public evidence yet that small-model agentic performance was bottlenecked on post-training data, not capacity — and that has a direct implication: the gap between open and closed models on agent tasks is narrower and more volatile than parameter counts suggest.
Should you switch your agent to V4 Flash 0731?
For high-volume, well-scoped agent work: yes, run the evaluation this week. For low-volume work on hard, ambiguous tasks: probably not, and the reason is not benchmark score.
Three things to check before you migrate:
- Concurrency ceiling. The published limit is 2,500 concurrent requests. That is generous for most teams and a hard wall for a few. Confirm your peak fan-out fits.
- Codex compatibility, but verify your harness. The 0731 release added Codex support, which makes it close to drop-in for existing agent scaffolds. "Close to" is doing work in that sentence — tool-call formatting edge cases are where migrations break.
- Data residency. DeepSeek's API is hosted in China. For a large share of enterprise buyers this is a hard blocker regardless of price or score, which is exactly the split we documented in Chinese AI models now powering up to 46% of enterprise API traffic.
That third point is the one that decides most real procurement conversations, and it is barely mentioned in benchmark coverage.
Calling DeepSeek V4 Flash 0731 from your code
The API is OpenAI-compatible, so migrating an existing client is a base-URL and model-name change. Here is a minimal request:
curl https://api.deepseek.com/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $DEEPSEEK_API_KEY" \
-d '{
"model": "deepseek-v4-flash-0731",
"messages": [
{"role": "system", "content": "You are a terminal agent. Prefer small, verifiable steps."},
{"role": "user", "content": "Find every TODO in ./src and summarise them by file."}
],
"stream": false
}'
The system prompt is deliberately the first message and deliberately stable — that is what makes cache hits possible. At $0.0028 per million cached input tokens versus $0.14 on a miss, a stable prefix is a 50x discount on the repeated portion of every call. Order your messages so the invariant part comes first, always.
A quick way to sanity-check whether a migration is worth it before you do the engineering:
# Rough monthly cost of an agent workload, list prices as of August 2026
def monthly_cost(in_tok_per_day, out_tok_per_day, in_price, out_price, days=30):
return days * (in_tok_per_day / 1e6 * in_price + out_tok_per_day / 1e6 * out_price)
flash = monthly_cost(8e6, 2e6, 0.14, 0.28) # DeepSeek V4 Flash 0731
sol = monthly_cost(8e6, 2e6, 5.00, 30.00) # GPT-5.6 Sol
print(f"Flash: ${flash:,.2f}/mo Sol: ${sol:,.2f}/mo ratio: {sol/flash:,.0f}x")
Run that with your own token counts before you argue about benchmark points. In most agent workloads the output-token price dominates, and that is where the 100x-plus spread lives.
One practical warning from running these migrations: cheap models fail differently, not less. A frontier model that gets confused tends to say so. A small model that gets confused tends to confidently produce a plausible-looking wrong patch. Budget for stronger verification — a test run, a diff review step, a second-model check — and spend part of what you saved on catching that failure mode.
The angle everyone is missing: this repriced the agent loop, not the chat box
Most coverage compares V4 Flash to other models on a per-token basis. That framing understates what happened.
Agentic workloads are token-hungry in a way chat is not. An agent re-reads its context on every step, re-plans, retries, and burns tokens on failed branches — a single non-trivial task can be 50 to 200 model calls. At $5/M input, teams design around that: they cap iterations, shrink context, and skip verification steps purely to control spend. At $0.14/M, those constraints stop binding.
That means the real effect is not "the same agent, cheaper." It is a different agent — one that can afford to double-check its own work, run three candidate solutions and pick the best, or re-read the whole repo instead of guessing from a summary. Cheap tokens do not just reduce the bill; they unlock strategies that were previously uneconomic. The teams that win from this release will be the ones that redesign their loop, not the ones that swap a base URL.
This is the same dynamic that made local inference interesting — see vLLM vs Ollama for the self-hosted version of the same argument.
Frequently Asked Questions
What is DeepSeek V4 Flash 0731? DeepSeek V4 Flash 0731 is the production release of DeepSeek's small agentic model, shipped July 31, 2026. It has 284 billion total parameters with 13 billion active, a 1M-token context window, and scores 82.7 on Terminal-Bench 2.1 at $0.14/$0.28 per million input/output tokens.
How much does DeepSeek V4 Flash cost? $0.14 per million input tokens on a cache miss, $0.0028 per million on a cache hit, and $0.28 per million output tokens. That is roughly 35x cheaper on input than GPT-5.6 Sol at $5.00/M.
Is DeepSeek V4 Flash better than DeepSeek V4 Pro? On agent benchmarks, yes. V4 Flash 0731 scores 82.7 on Terminal-Bench 2.1 against 72.1 for the V4 Pro preview, despite being roughly one-fifth the size. Pro retains advantages on some knowledge-heavy tasks, but for terminal-driven agent work the smaller model now leads.
Does DeepSeek V4 Flash 0731 work with Codex? Yes. The 0731 release added Codex compatibility, making it close to a drop-in replacement in existing agent harnesses. Test your tool-call formatting before migrating a production pipeline.
What is Terminal-Bench 2.1? Terminal-Bench is an agentic benchmark that runs a model inside a real terminal against multi-step engineering tasks, scoring whether the task was actually completed. It measures error recovery and tool use rather than recall, which is why post-training improvements move it so sharply.
Is DeepSeek V4 Flash open weights? DeepSeek has historically released open weights for its models, but the 0731 checkpoint is primarily distributed through its API. For a genuinely open-weight frontier alternative in 2026, look at Thinking Machines' Inkling, released under Apache 2.0.
The verdict
DeepSeek V4 Flash 0731 is the best price-performance model available for agentic coding in August 2026, and it is not particularly close. Two points behind Claude Opus 4.8 on Terminal-Bench while costing a rounding error of frontier pricing is the kind of result that resets a market segment rather than nudging it.
Our recommendation: if your agent workload is high-volume and your data can legally leave your jurisdiction, evaluate it this week and expect to redesign your loop around cheap tokens rather than just swapping the endpoint. If you are in regulated enterprise, keep watching — the score matters even if you cannot buy it, because it sets the price ceiling everyone else now has to argue against.
Want the frontier comparison alongside this? Read Claude Opus 5 explained for what the top of the market still buys you.
The old rule was that you paid for intelligence. The new rule is that you pay for the last three points — and DeepSeek just made everyone justify the invoice.
