DeepSeek Says Its New Coding Model Scores 42.3% on Terminal-Bench 4.0. Vals.ai Measures 31.82%.
DeepSeek’s launch table puts V4.1-Flash-1017 at 42.3% on Terminal-Bench 4.0. Vals.ai ran the same 66 tasks and measured 31.82%, yet DeepSeek’s scores for Claude Opus 5, GPT-5.6 Sol and GLM 5.3 land within 2.02 points of Vals’ own.
TL;DR
- 📉 10.48 points apart: DeepSeek reports 42.3% on Terminal-Bench 4.0. Vals.ai’s independent leaderboard lists 31.82%, rank 21 of 58.
- 🎯 Only DeepSeek’s own rows drift: DeepSeek’s numbers for Opus 5, GPT-5.6 Sol and GLM 5.3 sit within 2.02 points of Vals’. Both DeepSeek models sit 10 to 12 points higher.
- 📈 The upgrade is real: on Vals, 1017 gains 12.12 points over V4.1-Flash for 4 cents more per test. DeepSeek’s own table shows an 11.1-point gain.
- 💸 Still the cheap pick: $0.54 a test and $1.70 per solved task, against $20.26 for Claude Opus 5.5. Server-side Search & Fetch billing starts November 1.
What DeepSeek shipped on October 17
DeepSeek released V4.1-Flash-1017 on October 17, 2026, as an update to V4.1-Flash. It now answers to the deepseek-flash model name in the API. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but DeepSeek has retired those models and routes their requests to 1017 at the Flash price.
The release note promises “stronger benchmark performance, faster task completion, and shorter outputs,” and calls it a Pareto improvement in performance and cost. The weights are on Hugging Face under the MIT license, where the safetensors metadata lists 763B parameters.
The numbers sit in a benchmark table image on the release page. DeepSeek compares 1017 with its own V4.1-Flash, Kimi K3, GLM 5.3, GPT-5.6 Sol and Claude Opus 5. Here are the coding rows.
| Benchmark (DeepSeek-reported) | V4.1-Flash-1017 | V4.1-Flash | Kimi K3 | GLM 5.3 | GPT-5.6 Sol | Claude Opus 5 |
|---|---|---|---|---|---|---|
| Terminal-Bench 4.0 | 42.3 | 31.2 | 12.6 | 37.9 | 39.9 | 51.8 |
| DeepSWE v1.1 | 77.0 | 74.2 | 67.5 | 66.9 | 73.0 | 74.0 |
| ProgramBench | 29.1 | 20.3 | 17.5 | 19.0 | 23.0 | 37.0 |
| NL2Repo-Bench | 68.2 | 65.4 | 58.0 | 58.0 | 56.8 | 75.3 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 | 46.7 | 48.8 | 45.8 | 50.3 |
Source: DeepSeek’s release note, 2026/10/17. DeepSeek’s footnote says Terminal-Bench 4.0 ran in the Claude Code framework and DeepSWE in the Mini-SWE-V2 framework.
Same 66 tasks, an independent run
Vals.ai runs Terminal-Bench 4.0, a set of 66 community-contributed tasks, for every model in the same harness: mini-swe-agent, with a single tool, bash, and no step or cost limit. Each score is an average of 3 runs. Its leaderboard, updated October 29, lists 1017 at 31.82%.
Now line up every model that appears in both DeepSeek’s table and Vals’ board.
| Model | DeepSeek’s table | Vals.ai | Gap (DeepSeek − Vals)* |
|---|---|---|---|
| DeepSeek V4.1-Flash-1017 | 42.3% | 31.82% | +10.48 |
| DeepSeek V4.1-Flash | 31.2% | 19.70% | +11.50 |
| GPT-5.6 Sol | 39.9% | 37.88% | +2.02 |
| GLM 5.3 | 37.9% | 38.89% | −0.99 |
| Claude Opus 5 | 51.8% | 53.53% | −1.73 |
| Kimi K3 | 12.6% | 17.17% | −4.57 |
*Our math, from the two published tables.
Terminal-Bench 4.0: DeepSeek’s number minus Vals.ai’s, in points
Sources: DeepSeek release note (2026/10/17); Vals.ai Terminal-Bench 4.0, October 29, 2026. Bar length relative to the largest gap, 11.50.
Three of the four rival models land within 2.02 points of Vals. Kimi K3 is off by 4.57, and in Kimi’s favor on Vals. Both DeepSeek models come out more than 10 points higher in DeepSeek’s table than on Vals.
Why the harness is the likely answer
The two runs use different agents. DeepSeek ran Terminal-Bench 4.0 inside Claude Code, a full agent harness with file editing, search and planning tools. Vals hands every model one bash tool. A model trained to drive a rich harness can lose ground when it only gets a shell.
DeepSeek also leans into that harness on purpose. Its API docs list Claude Code and Codex as supported agent integrations, and its new server-side Search & Fetch tools are wired into DeepSeek Harness, Claude Code and Codex. If you run 1017 inside Claude Code, DeepSeek’s number is closer to your setup than Vals’ number.
What the harness story doesn’t explain is the pattern. If Claude Code lifted every model, the rivals in DeepSeek’s table should drift up too. They don’t: 3 of the 4 match the bash-only scores within about 2 points. DeepSeek doesn’t say whether it re-ran the rivals or copied their scores from elsewhere, so the table can’t tell you which explanation is right.
The rule that falls out of this: a vendor’s own chart tells you the direction of a change. An independent board with one harness for everyone tells you where a model ranks.
The improvement is real
Here the two sources agree. DeepSeek’s table shows 1017 gaining 11.1 points over V4.1-Flash on Terminal-Bench 4.0. Vals measures a 12.12-point gain, from 19.70% to 31.82%.
| Vals.ai, Terminal-Bench 4.0 | V4.1-Flash | V4.1-Flash-1017 | Change |
|---|---|---|---|
| Accuracy | 19.70% | 31.82% | +12.12 points |
| Cost per test | $0.50 | $0.54 | +8% |
| Tokens per test | 31.53M | 39.65M | +26% |
| Duration per test | 48m22s | 41m10s | 7m12s faster |
The “faster” claim holds on Vals. The “shorter outputs” claim is harder to check: Vals counts input and output tokens together, and the total went up 26%. Agents that survive longer on a task reread more context, so a higher total can still hide shorter answers.
The price math still wins
Even at 31.82%, 1017 is one of the cheapest ways to solve a Terminal-Bench task. The table below divides Vals’ cost per test by accuracy.
| Model | Accuracy | Cost / test | Tokens / test | Cost per solved task* |
|---|---|---|---|---|
| Claude Opus 5.5 | 65.15% | $13.20 | 28.27M | $20.26 |
| Kimi K3.1 Pro | 63.64% | $4.57 | 16.57M | $7.18 |
| GPT-6.1 Sol | 55.05% | $1.72 | 2.86M | $3.12 |
| GLM 5.3 Flash | 37.37% | $1.22 | 35.66M | $3.26 |
| Claude Haiku 5.5 | 36.36% | $3.12 | 45.50M | $8.58 |
| DeepSeek V4.1-Flash-1017 | 31.82% | $0.54 | 39.65M | $1.70 |
| MiMo V2.6 Pro | 31.31% | $0.46 | 26.60M | $1.47 |
*Our math: cost per test divided by accuracy. Every other number: Vals.ai Terminal-Bench 4.0, updated October 29, 2026.
Terminal-Bench 4.0, cost per solved task (USD)
Source: Vals.ai, October 29, 2026; our division. Bar length relative to Opus 5.5’s $20.26.
Opus 5.5 costs 24x more per test and solves about 2x as many tasks, so it comes out roughly 12x more expensive per solved task. 1017 burns 13.9x the tokens of GPT-6.1 Sol and still costs 31% of Sol per test. That is what DeepSeek’s list price buys you.
| Per 1M tokens | deepseek-flash, peak | deepseek-flash, off-peak | Claude Opus 5.5 |
|---|---|---|---|
| Input (cache miss) | $0.30 | $0.15 | $4.00 |
| Input (cache hit) | $0.006 | $0.003 | $0.20 |
| Output | $1.20 | $0.60 | $20.00 |
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. Everything else, including weekends, bills at half. Vals doesn’t say which rate it used, so if you run agents off-peak, your cost per task can be lower than the table.
One nuance: MiMo V2.6 Pro solves tasks a little cheaper, at $1.47, but takes 2h42m per test to 1017’s 41m10s. If your agent runs while you wait, 1017 is the cheap model that finishes.
The November 1 change
DeepSeek opened its own Search and Fetch APIs to everyone, at $0.75 per 1,000 successful requests each. When you use DeepSeek through DeepSeek Harness, Claude Code or Codex, server-side tool calls already hit those APIs, and so far they have been free.
That ends at 00:00 Beijing time on November 1, 2026, which is 16:00 UTC on October 31. If your agent searches the web a lot, add a line for it in your budget. At $0.75 per 1,000, a run that makes 200 search calls adds $0.15, close to a third of a full Terminal-Bench test.
Fun facts
- On DeepSeek’s own DeepSWE v1.1 table, 1017 scores 77.0, ahead of Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0. Its cost chart labels that point “max · 77.0%” at the far left of a log-scale cost axis.
- DeepSeek’s footnote marks two HLE numbers for 1017: 42.6 on the full set and 44.0 on the text subset.
- On Automation-Bench v1.0.6, DeepSeek reports 1017 at 60.5, ahead of every rival in its table, including Claude Opus 5 at 50.3.
- DeepSeek recommends DeepSeek Harness v0.2.1-rc.1 or later to run the new model.
What this means if you code with AI
Read launch charts for the delta, not the rank. DeepSeek’s table got the size of the 1017 upgrade right, within about 1 point. It got the level wrong by about 10. Use an independent board with one harness for everyone when you rank models against each other.
Your harness is part of the model. 1017 is tuned for Claude Code, Codex and DeepSeek’s own harness. Test it in the agent you actually use before you trust either number.
Use it where volume beats precision. At $1.70 per solved task, 1017 makes sense for subagents, test fixing, refactors you can retry, and anything you would run 10 times. For a hard one-shot change, Opus 5.5 or Kimi K3.1 Pro still solves about twice as many tasks.
Watch the new line item. Server-side Search & Fetch stops being free on November 1. Check how often your agent searches.
FAQ
What is DeepSeek-V4.1-Flash-1017?
An update to DeepSeek V4.1-Flash released on October 17, 2026. It now serves the deepseek-flash model name in DeepSeek’s API, and its weights are on Hugging Face under the MIT license. DeepSeek says it is stronger on benchmarks, faster and produces shorter outputs than the previous version.
What does DeepSeek-V4.1-Flash-1017 score on Terminal-Bench 4.0?
DeepSeek reports 42.3%, run in the Claude Code framework. Vals.ai’s independent leaderboard, which runs every model in the mini-swe-agent harness and averages 3 runs, lists 31.82% at $0.54 per test, rank 21 of 58.
Why are the two Terminal-Bench scores different?
The harnesses differ: DeepSeek’s footnote says its Terminal-Bench 4.0 tests used the Claude Code framework, and Vals gives every model one tool, bash, in mini-swe-agent. That is the most likely cause. What harness alone doesn’t explain is why DeepSeek’s published numbers for Claude Opus 5, GPT-5.6 Sol and GLM 5.3 sit within 2.02 points of Vals’ while both DeepSeek models sit 10 to 12 points higher.
How much does DeepSeek-V4.1-Flash-1017 cost?
Through the deepseek-flash API name: $0.30 per million input tokens (cache miss), $0.006 on a cache hit and $1.20 per million output tokens at peak, and half that off-peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday.
Is DeepSeek-V4.1-Flash-1017 good for coding agents?
On cost, yes: Vals.ai measures $1.70 per solved Terminal-Bench 4.0 task, against $3.12 for GPT-6.1 Sol and $20.26 for Claude Opus 5.5. On raw success rate it solves about half as many tasks as Opus 5.5, so it fits high-volume or retry-friendly work better than hard one-shot tasks.
What changes for DeepSeek users on November 1?
Search and Fetch calls made through server-side tools in DeepSeek Harness, Claude Code and Codex stop being free. DeepSeek bills them at $0.75 per 1,000 successful requests from 00:00 Beijing time on November 1, 2026, which is 16:00 UTC on October 31.
Where MegaBrain fits
Full disclosure: we sell an AI coding agent, and we did not run these benchmarks. MegaBrain’s terminal coding agent puts many models behind one plan: free models on Free, included models such as GLM, Kimi, MiniMax, DeepSeek, Qwen and Grok on the paid plans, Claude, GPT and Gemini from Pro, plus your own API keys and local models, and Auto picks one for each task. See the plans. MegaBrain is not affiliated with DeepSeek, Vals.ai or Anthropic.
Sources
All pages opened October 30, 2026.
- DeepSeek API Docs — DeepSeek-V4.1-Flash-1017 release note (2026/10/17), with benchmark table and DeepSWE cost chart
- DeepSeek API Docs — Models & Pricing (peak and off-peak rates, Search & Fetch billing)
- DeepSeek API Docs — Your First API Call (deepseek-flash model name, retired legacy names)
- Hugging Face — deepseek-ai/DeepSeek-V4.1-Flash-1017 (MIT license, weights)
- Vals.ai — Terminal-Bench 4.0 leaderboard (updated October 29, 2026)
- Anthropic — Claude API pricing (Opus 5.5, Opus 5)
MegaBrain · free to start
Try a free AI coding agent in your terminal or your editor.
No card: the free models, your own API keys and local models with no limits from us. 90+ models on one plan from $8 a month.
npm install -g @getmegabrain/cliThen run megabrain in your project and sign in once.
Newsletter
Stay in the loop
Get our latest AI research and guides. No spam, unsubscribe anytime.