DeepSeek V4-Flash Jumped 25 Points on Terminal-Bench Without Getting Bigger

DeepSeek V4-Flash Jumped 25 Points on Terminal-Bench Without Getting Bigger

6 min readAugust 1, 2026

Quick verdict

DeepSeek pushed a public-beta update to its V4-Flash API that behaves like a new model without being one. Same 284B total, 13B active parameters, same 1M context, but the agent benchmarks moved a lot. Terminal-Bench went from 56.9 to 82.7, and DeepSeek released the open weights under MIT on Hugging Face the same day. On Artificial Analysis it sits one point behind GPT-5.6 Luna while costing about 60 percent less per task on DeepSeek's own API. If you run coding agents, this is the cheapest thing on the frontier right now that you can also download and host yourself.

What actually shipped

The headline is that the jump came from post-training, not scaling. DeepSeek did not add parameters or change the architecture. It reworked the training after pretraining and got a much stronger agent out of the same weights.

  • Terminal-Bench 2.1 climbed to 82.7, up 25.8 points from the April preview's 56.9. On DeepSeek's own table it now beats V4-Pro-Preview across the coding and agent benchmarks it lists.
  • On the Artificial Analysis Intelligence Index the model rose from 40 to 50, landing one point behind GPT-5.6 Luna at 51 and roughly tying Gemini 3 Flash.
  • Pricing is $0.14 per million input tokens and $0.28 per million output, with a 98 percent cache-hit discount that drops cached tokens to $0.0028 per million.
  • DeepSWE sits at 54 percent PASS@1, which DeepSeek reports as level with Claude Sonnet 5 and Grok 4.5, though the leaderboard has not independently verified it yet.
  • The API now speaks the Responses API format and DeepSeek says it is fully adapted for Codex.

Open weights landed almost immediately, which is the part local users cared about most. The vLLM team pointed out the serving shape: 256 routed experts with 6 active per token, 1M context, three reasoning-effort levels, and a DSpark speculative decoding module you can turn on with a single flag. Unsloth had runnable quants out fast, needing roughly 168GB of RAM for a lossless 4-bit build and about 110GB for 3-bit. That is heavy for a laptop but reachable on a well-specced workstation, which is the whole point of an open release at this size.

One caveat worth keeping in view: these are benchmark and vendor numbers from launch day, and some are DeepSeek's own charts. Artificial Analysis had to correct a cache-hit display bug in the first hours, and DeepSWE parity is still unverified. The direction is clear, but treat the exact figures as provisional until third parties finish testing.

Why it matters

The timing is the story. A day earlier OpenAI cut GPT-5.6 Luna by 80 percent and Terra by 20 percent, and a lot of people read this Flash update as the direct answer. Now the two cheapest points on the frontier are trading blows on price, except one of them ships open weights you can run without asking anyone. That is a different kind of pressure than a closed price cut, because it puts a floor under what anyone can charge for this tier of intelligence.

For coding-agent workflows the practical shift is that Flash is good enough to be a default, not a fallback. Cline made the updated model free in its tool, developers wired it into Codex through routers that keep GPT, Grok, and Kimi in the same picker, and someone stood up a free public endpoint within hours. When the cheap model gets close enough to the expensive one, your bill stops being about which lab you trust and starts being about which tier each task actually needs. That routing question is the one that saves real money, and it is the case we walk through in cutting AI coding agent costs with model routing.

The bigger pattern is that open-weight models keep closing the gap on the paid frontier faster than the price of paid frontier plans comes down. A 284B model that runs on a workstation and ties Sonnet 5 on a coding benchmark would have been a headline a year ago. Now it is a Thursday. We track where that leaves the open ecosystem in open-source AI models in 2026, and what the falling token prices mean for anyone paying a monthly bill in where AI pricing is heading. If you would rather not pin yourself to one lab at all, running several models behind one subscription is covered in the best app for running multiple AI models.

Video: DeepSeek V4-Flash and the open-weight price war

This covers the V4-Flash update, the benchmark jump, and how it lands against the GPT-5.6 price cuts from the day before.

FAQ

Is DeepSeek V4-Flash actually open weights?

Yes. DeepSeek posted the weights on Hugging Face under an MIT license the same day it updated the API, so you can download, run, quantize, and fine-tune it with few restrictions. That is looser than most open-weight licenses and part of why the release got so much attention.

How big a machine do I need to run it?

It is a 284B total, 13B active mixture-of-experts model, so you do not load all the weights at once, but you still need the RAM to hold them. Unsloth's quants need roughly 168GB of RAM for a lossless 4-bit build and about 110GB for 3-bit. That rules out most laptops and points at a workstation or server.

How does it compare to GPT-5.6 Luna?

On the Artificial Analysis index it is one point behind Luna, 50 to 51, while costing around 60 percent less per task on DeepSeek's own API. The short version is near-parity on the benchmark at a much lower price, with the added option of self-hosting. For the other side of that price war, see our writeup of OpenAI's 80 percent Luna cut.

Should I switch my coding agent to it?

It is worth testing on your own tasks rather than trusting the launch charts. It is already free in Cline and easy to route into Codex, so the cost of trying it is low. The tier that fits depends on your work, which is the point of comparing agents rather than picking one, as we do in the best AI coding agents guide.

Sources

Further reading

Try all the models mentioned in this article

Admix gives you GPT-5, Claude, Gemini, and 350+ AI models in one app. Compare responses side by side. Free to start.

Start free on Admix

Related articles