Kingy AI
Subscribe, 2.1M subscribersWork with Kingy
AI News

DeepSeek V4-Flash 0731 Is Live in Codex, Claiming 82.7 on Terminal-Bench

July 31, 2026 AI News
DeepSeek V4-Flash-0731 company-reported agent benchmark scores

DeepSeek pushed a post-trained V4-Flash update into public beta on July 31, 2026 with native Responses API support for Codex clients. Update, checked October 10: DeepSeek now routes the legacy deepseek-v4-flash API name to V4.1-Flash, and V4-Pro has also launched. The 0731 benchmark and price figures below are launch-era evidence, not current API terms.

Launch Tracker: Read the verified DeepSeek V4-Flash-0731 launch record for the source checklist, pricing snapshot and availability details.

At the July 31 launch, DeepSeek identified the release as DeepSeek-V4-Flash-0731 and described it as retaining the architecture and parameter count of April’s V4-Flash preview. The company said the change was in post-training and reported much stronger agent and coding results, led by an 82.7 score on Terminal-Bench 2.1.

Cline quickly drew attention to the launch claim. The July 31 article recorded the April preview at 56.9, making the reported result a 25.8-point increase. That comparison remains an article-recorded launch snapshot rather than a newly recertified vendor figure. DeepSeek’s current Hugging Face model card, checked October 10, instead shows the preview at 61.8, a 20.9-point gap from 82.7. The more practical news for developers is that the updated model can now speak the protocol Codex uses.

What DeepSeek released

DeepSeek announced the official V4-Flash API as a public beta on July 31, 2026. The company’s changelog says the API call remains the same: developers select deepseek-v4-flash to reach the latest version.

  • Release: DeepSeek-V4-Flash-0731 public API beta
  • What changed at launch: DeepSeek described post-training changes, with the architecture and model size unchanged
  • Availability: DeepSeek API and MIT-licensed weights on Hugging Face
  • Codex support: native Responses API compatibility
  • Context: one million tokens, with a listed maximum output of 384,000 tokens
  • July 31 article pricing snapshot: $0.14 per million uncached input tokens and $0.28 per million output tokens

At the July 31 announcement, DeepSeek said the V4-Pro API and its app and web models were unchanged. The 0731 checkpoint itself is no longer limited to DeepSeek’s hosted API. Kimmonismus highlighted the weight release, and DeepSeek’s official DeepSeek-V4-Flash-0731 repository confirms a public MIT-licensed release. The July 31 article recorded 48 Safetensors weight shards and a 284-billion-parameter mixture-of-experts architecture with 13 billion parameters active per token. Those figures remain a dated launch snapshot: the current Hugging Face page, checked October 10, labels the repository “304B params” and says a speculative-decoding module is attached, while the retained current page does not independently substantiate the launch-era shard count or exact architecture figures.

The benchmark jump is broad, but still a vendor claim

DeepSeek’s release table compares V4-Flash-0731 with both the earlier V4-Flash preview and V4-Pro preview. The updated Flash model leads both previews across every agent benchmark shown.

Selected July 31 article-recorded company results (V4-Flash-0731 vs V4-Flash Preview):

  • Terminal-Bench 2.1: 82.7 vs 56.9, up 25.8 points
  • NL2Repo: 54.2 vs 39.4, up 14.8 points
  • Cybergym: 76.7 vs 38.7, up 38.0 points
  • DeepSWE: 54.4 vs 7.3, up 47.1 points
  • Toolathlon-Verified: 70.3 vs 49.7, up 20.6 points
  • Agents’ Last Exam: 25.2 vs 15.8, up 9.4 points

On October 10, DeepSeek’s current Hugging Face model card still showed the same five result/preview pairs below Terminal-Bench, but it listed the Terminal-Bench 2.1 preview at 61.8 rather than 56.9. Against the unchanged 82.7 result, that current-page comparison is a 20.9-point gap. The list above preserves the July 31 snapshot rather than silently rewriting the launch record.

Those gains are unusually large for a release that leaves the base architecture untouched. They also need context. DeepSeek ran the public code-agent evaluations with its forthcoming “DeepSeek Harness” in minimal mode, using maximum reasoning effort, top_p=0.95 and temperature=1.0. At the July 31 announcement, the harness had not yet been released. DeepSeek now has a public Harness repository, but Kingy.ai has not reproduced the original benchmark setup or its scores.

Two other scores in the announcement, DSBench-FullStack at 68.7 and DSBench-Hard at 59.6, come from DeepSeek’s internal test sets. They are useful as company-reported indicators, not independent evidence. Kingy.ai has not rerun these benchmarks or tested V4-Flash-0731 in a production repository.

Why the Codex support matters

Codex clients use OpenAI’s Responses API rather than the older Chat Completions format. V4-Flash-0731 now supports that interface natively, so developers can add DeepSeek as a model provider instead of routing requests through a compatibility layer.

DeepSeek’s Codex integration guide provides an automated setup script and a manual configuration path. Both create a custom model catalogue and add a DeepSeek provider to Codex’s shared ~/.codex/config.toml file. Because Codex CLI, the desktop app and the VS Code extension read that same file, one configuration makes the model available across all three clients.

At the July 31 launch, only V4-Flash worked with Codex. DeepSeek released V4-Pro on August 13, 2026, and its current Codex guide includes both models. In the ChatGPT desktop app, locally configured models appear under the generic “Custom” label, so users should verify the selected provider and model in their configuration rather than relying on the picker label alone.

The price makes the result hard to ignore

The July 31 article recorded V4-Flash pricing at $0.0028 per million cached input tokens, $0.14 per million uncached input tokens and $0.28 per million output tokens. Its current pricing page, checked October 10, instead lists V4.1-Flash as deepseek-flash and routes legacy deepseek-v4-flash requests to it. Per million tokens, V4.1-Flash is $0.003 cached input, $0.15 uncached input and $0.60 output off-peak; peak rates are $0.006, $0.30 and $1.20. The current page lists a one-million-token context window and support for thinking and non-thinking modes.

That combination makes V4-Flash-0731 an obvious candidate for cost-sensitive agent workloads: long repository context, native tool-oriented API support and a low token price. It does not prove the model will outperform a more expensive competitor on a particular codebase. Terminal agents are sensitive to scaffolding, prompts, tool permissions, latency limits and the exact task mix. A team choosing a coding model should rerun its own repository-level evals before moving production work.

DeepSeek’s current pricing page, checked October 10, lists peak and off-peak rates, specifies the peak windows and says rates may change; readers should check it before estimating a workload.

What to watch next

The first question is reproducibility. DeepSeek’s now-public Harness gives independent evaluators the code needed to examine how much of the reported gain comes from the model and how much comes from the agent scaffold, but Kingy.ai has not reproduced the July 31 setup or scores. The second is practical self-hosting. The 0731 weights are available under MIT, but Kingy.ai has not tested a local deployment or verified parity with DeepSeek’s hosted API. The launch-era architecture description still implies a substantial hardware and serving project.

The third was V4-Pro. That July 31 expectation is now resolved: DeepSeek released V4-Pro on August 13, and the current Codex guide covers both Flash and Pro. The 0731 checkpoint remains the subject of the benchmark table above; current API traffic under the legacy Flash name reaches V4.1-Flash.

For developers, the sensible next step is a controlled trial: configure V4-Flash through DeepSeek’s official guide, run a representative set of repository tasks, and compare completion quality, tool reliability, latency and total token cost against the model already in use.