A $0.20-per-Task Update That Beats Opus 5 on Code

On October 17, 2026, DeepSeek pushed V4.1-Flash-1017 as a minor revision to V4.1-Flash. The headline number: 77.0 on DeepSWE v1.1, measured by DeepSeek on its own mini_swe_v2 harness. For context, that sits above Claude Opus 5’s 74.0 and GPT-5.6 Sol’s 73.0 on the same benchmark as reported by DeepSeek and compiled by Xenospectrum.

The model inherits V4.1-Flash’s architecture: 552 billion total parameters with 8 billion active per token, a 1 million-token context window, and a 384K maximum output per request, according to SmartScope’s technical breakdown. Pricing is unchanged at $0.15 input / $0.60 output per million tokens off-peak, and $0.30 / $1.20 at peak. MIT-licensed weights — roughly 763 billion parameters in safetensor format — are live on Hugging Face and ModelScope.

The Benchmark Numbers — And Their Caveats

The table below compiles scores from DeepSeek’s model card as reported by Xenospectrum and SmartScope. DeepSeek ran every figure — none have been independently replicated.

Three caveats matter here. First, all scores come from DeepSeek’s own model card — the standard caveat for any vendor-run benchmark. Second, the comparison models’ scores were also evaluated by DeepSeek; independent evaluators may produce different numbers. Third, Xenospectrum notes that the Terminal-Bench harness changed between the V4.1-Flash and the 1017 update: the new run used the Claude Code harness, while the previous version used the DeepSeek Harness. The apparent 11-point jump (31.2 → 42.3) is therefore not a clean apples-to-apples comparison. Run your own evals before switching.

The Cost Math: $0.20 vs $0.65 per Completed Task

Benchmark percentages tell you what a model can do; cost-per-task tells you what it costs to actually do it. Xenospectrum’s cost-performance chart — reading DeepSeek’s vendor-produced mini_swe_v2 data — puts DeepSeek V4.1-Flash-1017 at approximately $0.20 per completed task at 77% success. GPT-6.1 Sol at high reasoning effort sits at roughly $0.65 per task at 75% success.

Sources: DeepSeek pricing from SmartScope; GPT-6.1 Sol pricing from Local AI Zone model tracker; Claude Opus 5.5 pricing from MetricNexus; Gemini 4 Argon pricing from Local AI Zone model tracker. Cost-per-task estimates are approximate readings from a vendor-produced cost-performance chart in the mini_swe_v2 environment — treat them as indicative, not benchmarked figures.

The headline claim: you get roughly the same task-completion rate as GPT-6.1 Sol at a third of the cost. That ratio holds only in DeepSeek’s own harness. Independent evaluations on production codebases typically widen the gap on harder tasks. For a fairer comparison against your specific stack, see the vortx.ch coding agent benchmark guide, which walks through replicating key benchmarks with your own test suite.

The Free Search API Window Closes October 31

DeepSeek launched two new tool-call APIs alongside the model update. The Search API and Fetch API are both priced at $0.75 per 1,000 requests. Through October 31, 2026 at 16:00 UTC, calls made through server-side tools in Claude Code, OpenAI Codex, or DeepSeek Harness v0.2.1-rc.1 or later are free, per SmartScope’s API breakdown. After that, the standard rate applies. Teams prototyping search-augmented agents have about 11 days from today to test at zero marginal cost.

Existing API integrations that route to deepseek-flash, deepseek-v4-flash, or deepseek-v4-flash-vision-exp will hit 1017 automatically. No code changes needed unless your system pins to an older checkpoint by name.

For Swiss & EU teams

DeepSeek’s first-party API routes data through servers in China, which creates a cross-border transfer under GDPR and the Swiss Federal Act on Data Protection (FADP). For teams that cannot use the first-party endpoint, MIT-licensed weights enable self-hosting inside the EU — though at 552 billion parameters the model requires hundreds of gigabytes of VRAM and is impractical on a single consumer GPU. Lyceum Technology hosts DeepSeek V4 Flash in the eu-north1 region and states that inference prompts are processed without retention; verify whether the 1017 checkpoint is available on their serverless catalogue and request a Data Processing Agreement before connecting production data. No Switzerland-specific region is available at this writing.

What to Do Now

If your pipeline already uses DeepSeek V4.1-Flash: the 1017 update is a drop-in; test your regression suite against the new checkpoint before routing production traffic. If you are costing out a switch from Claude Opus 5 or GPT-6.1 Sol for high-volume agentic workloads, the $0.20/task estimate warrants a controlled evaluation — set up a 200–500-task sample from your own backlog and compare task success and output quality, not just benchmark percentages. The free Search & Fetch API window through October 31 makes this a low-cost week to prototype search-augmented agents.

Mistral Large 4’s MIT weights are due by end of October; a second open-weight frontier model at comparable pricing would give teams another serious option. The sub-cent token math that shaped October’s pricing landscape will get another stress test when those weights land.

Further Reading

• DeepSeek V4.1-Flash-1017: Changes, Benchmark Comparisons, API Pricing (SmartScope) — the most thorough technical breakdown of the release, including the full benchmark table and API token pricing.
• DeepSeek V4.1 Flash Update Tops Claude Opus 5 on Code-Fixing Benchmark (Xenospectrum) — the model comparison table and the cost-per-task chart this article draws on.
• How to Evaluate a New Frontier Model Before Switching (vortx.ch) — a step-by-step guide to running your own evals before committing budget to a new model.

Your turn: Is DeepSeek V4.1-Flash already in your production pipeline, or does the data-residency question keep it off the table? Reply to our newsletter or send us a note — we feature the best answers in the Friday Scorecard.

How this article was made: AI researched and wrote this article under the Adrian persona, using the sources linked above, and it was published automatically without a human edit. Editorial guidelines are set by Adi. Spotted an error? Tell us and we will correct it.