7.5 Times Smarter for the Same Price

On July 31, 2026, DeepSeek released the general availability of V4 Flash. The model has 284 billion total parameters and activates 13 billion per token. It costs fourteen cents per million input tokens and twenty-eight cents per million output tokens. And it just beat GLM-5.2, a model with 40 billion active parameters, on every single agent benchmark they share.

Let me say that again, because the numbers are the story and they deserve the repetition. A model with roughly one third the active parameter count of its closest rival outperformed that rival across eight independent agent benchmarks. The price did not change. The architecture did not change. The model did not get bigger.

It got smarter.

The numbers

DeepSWE is the headline. It is a software engineering benchmark that tests whether a model can autonomously resolve real GitHub issues, and the previous V4 Flash Preview scored 7.3 on it. The GA release scores 54.4. That is a 7.5 times improvement, and it came entirely from post-training: reinforcement learning, instruction tuning, whatever DeepSeek is doing in the gap between a base model and a shipped one. The weights are the same shape. The intelligence is not.

The rest of the agent suite tells the same story. Cybergym went from 38.7 to 76.7. AutomationBench went from 10.8 to 25.1. TerminalBench 2.1 landed at 82.7. NL2Repo reached 54.2. Toolathlon verified hit 70.3. DSBench-Hard, which is DeepSeek’s internal coding agent torture chamber, scored 59.6.

On the Artificial Analysis Intelligence Index, which aggregates performance across a broad suite of tasks, V4 Flash Preview sat at 40. The GA release sits at 50. That is tied with Gemini 3.6 Flash, which is a Google model running on Google infrastructure with Google’s entire optimisation stack behind it. DeepSeek matched it with an open-weight model that costs less than a cup of chai per million tokens.

What the independent numbers say

Artificial Analysis published their evaluation within hours of the release. Their Intelligence Index puts V4 Flash 0731 at 50, a 10-point jump from the April preview and, remarkably, 6 points ahead of V4 Pro, the larger sibling that has not yet received this post-training update. The model lands on their Pareto frontier for intelligence versus cost per task, meaning nothing cheaper scores higher and nothing that scores higher costs less.

The GDPval-AA v2 Elo, which measures performance on real-world agentic work tasks, jumped from 1189 to 1559. That is a 370-point gain in a system where 50-point moves are significant. Terminal-Bench 2.1 rose 17 points to 79 percent. The model improved on every single evaluation in the Intelligence Index without exception.

Two findings from the AA report deserve specific attention. First, V4 Flash 0731 used 12 percent fewer output tokens than its predecessor to complete the same benchmark suite. It is not just producing better answers; it is producing them more efficiently. Second, the AA-Omniscience improvement, a measure of how honestly a model communicates its own uncertainty, improved by 7 points purely through reduced hallucination. Accuracy stayed flat. The model did not get more confident; it started making fewer things up. Its hallucination rate of 84 percent is comparable to GPT-5.6 Terra and Mistral Medium 3.5, both of which are substantially larger and more expensive.

The cost comparison is where the story tilts from impressive to uncomfortable for competitors. OpenAI cut the price of GPT-5.6 Luna by 80 percent today, presumably in response to exactly this kind of pressure. Even after that cut, DeepSeek V4 Flash on DeepSeek’s first-party API is still approximately 60 percent cheaper per task than GPT-5.6 Luna max, a model that scores one point higher on the Intelligence Index. DeepSeek achieves this partly through a 98 percent cache hit discount on its own API, which is significantly more aggressive than the 90 percent discount most providers offer.

The model sits one point behind GLM-5.2 max on the Intelligence Index. GLM-5.2 activates 40 billion parameters per token. V4 Flash activates 13 billion. The gap between them is one point.

What did not change

The architecture is identical to the preview. Same 284 billion total parameters, same 13 billion active per forward pass, same Mixture of Experts routing. Same Hybrid Attention Architecture with Compressed Sparse Attention and Heavily Compressed Attention interleaved, which is the trick that makes the 1 million token context window actually work instead of quietly ignoring everything past 128K. Same MIT licence, same 160 gigabyte download on Hugging Face.

The pricing is identical. Fourteen cents per million input tokens. Twenty-eight cents per million output tokens. For context, GLM-5.2 costs more than that to run, and someone on X pointed out that you could load two instances of V4 Flash in parallel for less than one GLM-5.2 inference pass and still come out ahead on benchmarks. That is the kind of arithmetic that gets procurement teams to switch providers before the end of the quarter.

The API interface is compatible with both OpenAI ChatCompletions and Anthropic formats. The GA release also natively supports the Responses API format and has been specifically adapted for Codex, which means the integration story for anyone running coding agents is already done.

The interesting question

When a model improves this dramatically without any change to its architecture, the interesting question is what DeepSeek found in their training pipeline. They have not published the details yet, and they may never publish them. But the shape of the improvement tells you something.

The gains are concentrated in agent tasks. TerminalBench, DeepSWE, Cybergym, Toolathlon: these are benchmarks that test whether a model can plan, use tools, recover from errors, and sustain a multi-step workflow without losing the thread. They are not testing knowledge retrieval or single-turn reasoning. They are testing the model’s ability to act as an autonomous worker.

Post-training is where behaviour gets shaped. A base model can be brilliant at next-token prediction and still be terrible at deciding when to call a function, how to parse an error, or when to abandon a failed approach and try something else. Those are learned behaviours, and they are learned during the phase that DeepSeek clearly invested in heavily.

DeepSeek noted in their changelog that the model was tested using their own DeepSeek Harness minimal mode with max effort level, top-p at 0.95, and temperature at 1.0. That is a specific configuration for a specific purpose: getting the model to explore broadly during agent tasks rather than converging too early. It suggests they have been running extensive agent evaluations and tuning for exactly this regime.

What is still missing

V4 Pro, the larger sibling at 1.6 trillion total parameters and 49 billion active, is still in preview. Its official GA release is pending. The Flash GA release does not upgrade the Pro API or the consumer app models. If you are using DeepSeek through chat.deepseek.com, you are not getting today’s improvements yet.

And the honest caveat on agent benchmarks is that they are still benchmarks. A model that scores 54.4 on DeepSWE can resolve a certain percentage of curated GitHub issues in a controlled harness. How it performs inside your codebase, with your tooling, your test suite, and your particular flavour of undocumented architectural decisions, is a different question. The numbers are real, and they are impressive, and they are also a ceiling estimate for production use rather than a floor.

What it costs to care

One hundred and sixty gigabytes on Hugging Face, MIT licensed. Fourteen cents per million tokens through the API. A 50 on the Artificial Analysis Intelligence Index, tied with the best small model Google ships. And a 7.5 times jump on DeepSWE that nobody predicted because the model did not change size; it changed how it uses what it has.

The DeepSeek Harness minimal mode that produced these benchmark numbers will be released soon. When it is, the configuration that V4 Flash was tuned against will be reproducible by anyone. That matters more than any single benchmark score, because it means the evaluation is verifiable rather than asserted.

Full weights are expected in the coming weeks. When they land, this becomes the second highest open-weights model on GDPval-AA v2, behind Kimi K3 and ahead of GLM-5.2. For a model that costs fourteen cents per million input tokens and uses 12 percent fewer tokens than its own previous generation to do the same work, that is a difficult number to argue with.

Seven dollars and twenty-eight cents. That is what one million output tokens costs through this API. The model behind those tokens just landed on the Pareto frontier, one point behind a model three times its active size, for less than a cup of chai. Someone at DeepSeek found something in post-training that works, and they are not charging extra for it.