Within about a week, three vendors have put out fast, inexpensive models aimed at agents and coding: Inception’s Mercury 3 (October 19), MiniMax’s M3.1 Flash (also October 19) and an updated DeepSeek V4.1-Flash-1017 (announced over the weekend). All three compete on speed and price rather than on being the single smartest model available. This article lines up what each vendor has published. We have not run these models ourselves, every benchmark below is a vendor-reported figure that has not been independently checked, and prices can change, so treat this as a map of the claims rather than a verdict.
The numbers side by side
| Mercury 3 (Inception) | M3.1 Flash (MiniMax) | V4.1-Flash-1017 (DeepSeek) | |
|---|---|---|---|
| Architecture | Diffusion language model | Sparse model, 428B parameters in total (per MiniMax’s developer guide) | Not detailed in the announcement we read |
| Context window | 1 million tokens | Prompts up to 1 million tokens | Not stated in the announcement we read |
| Input price per 1M tokens | $0.20 list, $0.10 at launch | $0.15 flat | $0.15 off-peak (per SiliconANGLE) |
| Output price per 1M tokens | $0.70 list, $0.35 at launch | $0.60 | Not stated in the announcement we read |
| Cached input per 1M tokens | $0.02 list, $0.01 at launch | $0.015 | Not stated in the announcement we read |
| Speed claim | 1,068 tokens per second on NVIDIA GPUs | Close to three times faster than M3; peak output above 200 tokens per second | “Faster task completion, shorter outputs” than the previous version |
| Weights | Not stated | Promised for a later date | Published on Hugging Face and ModelScope |
| Headline benchmark | Vendor says it beats GPT-6 Luna and Claude Haiku 5.5 on a range of benchmarks | OSWorld V2 61.7; DeepSWE v1.1 73.7%; CyberGym 82% | DeepSWE v1.1 77% (per SiliconANGLE) |
Two things stand out. First, the prices are within a few cents of each other, so the choice is unlikely to be decided on list price alone. Second, the openness differs: DeepSeek has released the weights, MiniMax says it will, and Inception has not said.
Mercury 3: speed from a different architecture
Mercury 3 is a diffusion LLM. Instead of producing text one token after another, diffusion language models refine many positions in parallel, which is how Inception reaches more than 1,000 tokens per second. Inception says the model is more than 40 times larger than Mercury 1 and that benchmark intelligence tripled between the two, and it argues this shows text diffusion scales much like autoregressive models do. The company also reported that, on 50 live Mind2Web navigation tasks inside its own web-agent harness, Mercury 3 completed 50% and outperformed GPT-6.1-Sol at the same token volume and step count. That is a small, self-run sample, so it is best read as a demonstration rather than a ranking. For background on how diffusion differs from other generators, see our explainer on how diffusion models generate images.
MiniMax M3.1 Flash: cheaper agent runs
MiniMax says agentic tasks cost 70% less to run than on M3, credits “more aggressive sparsity at inference time” for the speed, keeps reasoning on for every request with five effort levels from Low to Max, and accepts text, image and video input. Its own testing put M3.1 Flash at 61.7 on OSWorld V2 (computer operation), 76.2 on BankerToolBench and 82% on CyberGym, a Berkeley-built benchmark on reproducing real software vulnerabilities. It is available through the MiniMax API, its M Plan subscription and the MiniMax Code desktop app. MiniMax also confirmed that “Space Bunny”, an anonymous model that appeared on OpenRouter in late September, was an early build, and said more than 200 trillion tokens passed through it. A more capable Pro model is promised.
DeepSeek V4.1-Flash-1017: the open-weight incumbent
DeepSeek describes the update as stronger on benchmarks, faster at finishing tasks and shorter in output than the previous version, a “Pareto improvement in performance and cost”. It highlights reinforcement learning with self-feedback for building complete projects from scratch, and recommends upgrading to DeepSeek Harness v0.2.1-rc.1 or later. Separately, DeepSeek opened its Search and Fetch APIs to everyone at $0.75 per 1,000 calls per endpoint; calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are free until billing starts on October 31, 2026 at 16:00 UTC. SiliconANGLE reports 77% on DeepSWE v1.1 versus 73.7% for MiniMax, and notes neither result has been independently verified.
One caution for organizations in the United States: on the same day, the Department of Homeland Security ordered DeepSeek products removed from federal civilian systems, which matters for anyone selling into government.
How to compare them without trusting the tables
Vendor benchmarks are chosen by the vendor, run on the vendor’s harness and rarely repeated by third parties within days. A 3-point gap on DeepSWE v1.1 between two self-reported results is well inside the noise you would expect from different scaffolding. A better approach is to build a small evaluation set from your own tasks, run all three, and measure cost per completed task rather than cost per token, because shorter outputs and fewer retries change the bill more than a cent in list price. Our guide on building an evaluation set for LLM apps walks through that process.
What to watch next
- Independent runs of DeepSWE v1.1, OSWorld V2 and similar tests on all three models.
- Whether MiniMax publishes the M3.1 Flash weights and what license applies.
- Whether Inception reveals more about the Mercury 3 architecture, size and training.
- The end of Mercury 3’s launch discount, which would raise its price per token.
- How DeepSeek’s US federal ban affects enterprise adoption outside government.
Key takeaways
- All three models target agent and coding workloads and cost roughly $0.10 to $0.20 per million input tokens.
- Mercury 3 trades the usual token-by-token method for diffusion to reach about 1,068 tokens per second, by Inception’s account.
- Only DeepSeek’s weights are public today; MiniMax’s are promised.
- Every benchmark here is vendor-reported and unverified; we did not test the models.
- Measure cost per finished task on your own data before switching.
Frequently asked questions
Which of the three is cheapest?
On launch pricing Mercury 3 is cheapest per million input tokens ($0.10, rising to $0.20 after the discount). MiniMax and DeepSeek list $0.15 for input. Output prices are only published for Mercury 3 and MiniMax in the material we reviewed.
Is a diffusion language model better than a standard one?
Not automatically. Diffusion can be much faster, but quality depends on the specific model and task. Inception’s claims about beating Claude Haiku 5.5 and GPT-6 Luna come from its own benchmark selection.
Can I run these models locally?
Only DeepSeek V4.1-Flash-1017 has published weights today. Whether a given machine can run it depends on the model size, which the announcement we read does not state.