Inception Mercury 3 vs MiniMax M3.1 Flash: The October 19 Speed Duel, Compared

On October 19, 2026, the AI industry got two new speed-first models in one day. Inception Labs released Mercury 3, a diffusion-based language model it says generates 1,068 tokens per second. MiniMax released M3.1 Flash, an agentic model at $0.15 per million input tokens. Both ship a 1-million-token context window, targeting the same workload: AI agents that make dozens of model calls per task, where latency and cost compound fast.

Both launched days after Anthropic's Claude Haiku 5.5 reset the small-model price floor — so every new release now answers the same question: why should a developer pay more? Here's what each model offers, what the numbers mean, and where each one's fine print hides.

Quick answer

• Inception's Mercury 3 claims 1,068 tokens per second at a launch price of $0.10 per million input tokens.
• MiniMax's M3.1 Flash charges a flat $0.15 per million input tokens and reports 73.7% on DeepSWE v1.1.
• Mercury 3 leads on vendor-reported speed; M3.1 Flash leads on vendor-reported agentic coding scores — and every benchmark in this story was run by the company selling the model.

What Mercury 3 actually is

Mercury 3 is the fourth generation of Inception's Mercury line, built on the architecture bet the company is known for: diffusion language modeling. Where conventional models write left to right, one token at a time, a diffusion model starts with a rough draft of the whole answer and refines it in parallel across denoising steps — the same coarse-to-fine idea behind image generators. Inception's argument: the speed advantage grows as models get bigger and "think" longer per answer.

The headline number is 1,068 tokens per second on NVIDIA GPUs — co-founder Volodymyr Kuleshov specified widely available NVIDIA Blackwell GPUs. Context is 1 million tokens, up from 260K on Mercury 2.5, and Inception says the model is more than 40x larger in total parameters than the original Mercury 1.

Availability is API-only: Inception's own API Platform, plus OpenRouter, Baseten, and Vercel. The API is OpenAI-compatible (model string mercury-3) with a reasoning_effort control. No open weights, none announced. Mercury 2 and 2.5 are closed to new users — existing customers with usage before October 19, 2026 keep access until November 15, 2026. See Inception's official models page for the pricing table.

What M3.1 Flash actually is

M3.1 Flash is MiniMax's efficiency-focused release: a mixture-of-experts model listed at 428B total parameters with roughly 23B active, built for agentic office work, coding, computer use, and cybersecurity. It takes text, image, and video inputs, returns text, and ships a 1M-token context window.

It arrived with a backstory. For weeks, an anonymous model called Space Bunny topped public leaderboards — MiniMax confirmed on launch day that Space Bunny was an early M3.1 Flash build, claiming it processed more than 200 trillion tokens in stealth. OpenRouter, which hosted the stealth run, reported Space Bunny as its #1 model for 10 straight days at 56.7 trillion tokens, with time to first token down 38% and throughput up 45% during the stealth period. OpenRouter brought it back on October 19 as M3.1 Flash and offered it free for 72 hours. Unite.AI's launch coverage has the full timeline.

Unlike Mercury 3, MiniMax says open weights are coming — promised for October 23, 2026 in the company's X post, though the launch blog only says "later," with the license still undisclosed. A Pro model is promised with no date.

Pricing: launch promo vs flat rates

Mercury 3 lists at $0.20 per million input tokens and $0.70 per million output — but launches 50% off: $0.10 and $0.35. Cached input is 90% off the input rate: $0.02 standard, $0.01 in the promotion. New accounts get 100 million free tokens. The catch: Inception has announced no end date for the discount. Budget for list price, not the promo — $0.20/$0.70 is the number that matters for production, because a promotion is not a price promise.

M3.1 Flash is simpler: $0.15 per million input tokens, $0.60 per million output, $0.015 per million cached input — flat, with no long-context surcharge. MiniMax also claims 70% lower usage cost on agentic tasks than M3 — but that comes from its own task-cost evaluations, not a list-price cut. Per token, Flash's rates are half of M3's standard $0.30/$1.20 (up to 512K input, doubling above it). MarkTechPost's breakdown confirms the flat pricing.

Under Mercury 3's promo, Inception undercuts MiniMax on input ($0.10 vs $0.15) and output ($0.35 vs $0.60); when the promo ends, positions flip ($0.20/$0.70 vs $0.15/$0.60). Cached input favors Mercury 3 either way ($0.01 promo vs $0.015) — which matters for agent loops reusing the same system prompt, where cache pricing often decides the real bill.

For context, the models both vendors benchmark against sit at $0.10 in / $0.50 out: GPT-6 Luna flat, and Claude Haiku 5.5 for prompts up to 100,000 tokens — rising to $0.50/$2.50 above that threshold, the pricing cliff we covered in our Claude Haiku 5.5 pricing breakdown.

Speed: read the footnotes before the headline

Mercury 3's 1,068 tokens per second is a generation-throughput figure — how fast the model emits tokens once running, measured by Inception on NVIDIA GPUs. Inception says that's 4–8x faster than the compared models, though its own pages disagree: the launch blog says 4–8x, the models page says 5x.

MiniMax claims output speeds "can reach" 200+ tokens per second in everyday use with a median (P50) above 170 — a peak-type number next to a median, not Inception's measurement. Artificial Analysis, an independent evaluator, measured 120 tokens per second with 1.08 seconds to first token — below MiniMax's claim, though test conditions differ.

The honest takeaway: Mercury 3 reports the higher number by a wide margin, but the companies measure different things on different hardware, and neither figure has been independently reproduced like-for-like. Throughput isn't user-perceived latency either — network delay, prompt processing, and tool calls all eat into it. Third-party data points remain modest: Octen AI measured a 0.9-second median generation time for Mercury 3.

Benchmarks: everyone's winning their own chart

Inception reports Mercury 3 beating GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 selected evaluations: SWE-Bench Atlas at 51.08 (18 points ahead of both), BrowseComp+ at 55.18 vs 50.50 for Luna and 45.78 for Haiku 5.5, IFBench at 76.84. It loses some: DeepSWE v1 goes to Haiku 5.5 (51.62 vs 46.31), and Haiku also leads the MRCR v2 comparison. Luna edges Mercury on one long-context test.

MiniMax reports 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 62.2 on Automation-Bench v1.0.6 (top of its own table), 82% on CyberGym, and 76.2 on BankerToolBench — ahead of DeepSeek V4.1 Flash on Automation-Bench and HLE (48% vs 43%) in its own comparison. The independent Artificial Analysis Intelligence Index scores it 42–43 — below DeepSeek V4.1 Flash 1017 (45), GLM-5.3 (45), and Xiaomi's MiMo-V2.6-Pro (46).

Watch the version skew: Mercury 3's coding number is DeepSWE v1 (46.31); M3.1 Flash's is DeepSWE v1.1 (73.7). Not the same test — don't read them as a head-to-head. The vendor grading its own homework is a starting point, not a verdict.

Where to actually get them

The fine print that matters

• Mercury 3's 50% launch discount has no announced end date. No guarantee it's still active when you read this — check Inception's models page before budgeting.
• M3.1 Flash's weights date is soft. October 23 comes from the company's X post; the launch blog says only "later." The undisclosed license matters for commercial self-hosting plans.
• M3.1 Flash's thinking controls have teeth. Thinking can't be switched off (a disabled flag returns HTTP 400, error code 2013), an unset effort field defaults to max, and pay-as-you-go request ceilings dropped from 500 to 200 per minute. Check your defaults — max thinking on every call is not the bill you planned.
• OpenRouter's 72-hour free window for M3.1 Flash started October 19. Reading this after October 22? It's over — check current listings.

Which one should you try first?

Mercury 3 — pros

• Highest reported throughput of any model compared here (1,068 tok/s).
• $0.10/$0.35 launch pricing with the cheapest cached-input rate in this comparison.
• OpenAI-compatible API across four providers, plus 100M free tokens for new accounts.

Mercury 3 — cons

• The promo price has no end date and no price promise behind it.
• API-only with no open weights planned; older Mercury models sunset November 15, 2026.
• Every benchmark is vendor-run, and the company's own pages disagree on the speed multiple (4–8x vs 5x).

M3.1 Flash — pros

• Flat, predictable pricing with no long-context surcharge — cheaper at list price once Mercury 3's promo ends.
• Strong vendor-reported agentic coding scores (73.7% DeepSWE v1.1, 62.2 Automation-Bench).
• OpenAI- and Anthropic-compatible access, multimodal inputs (text, image, video), weights promised October 23.

M3.1 Flash — cons

• Speed is measured differently and the claimed number is far lower than Mercury 3's; independent measurement (120 tok/s) trails the claim.
• Weights aren't out yet and the license is unknown.
• Artificial Analysis Index of 42–43 trails rival cheap models like DeepSeek V4.1 Flash (45).

FAQ

Is Mercury 3 faster than MiniMax M3.1 Flash?

On the vendors' own numbers — 1,068 vs a claimed 200+ tokens/sec — yes, by a wide margin. But the numbers measure different things on different hardware, and neither has been independently reproduced head-to-head.

Which is cheaper per token?

At Mercury 3's launch promo ($0.10 in / $0.35 out), Inception. At its standard price ($0.20/$0.70), MiniMax's flat $0.15/$0.60. The promo has no announced end date, so budget at list price.

Can I self-host either model?

Not Mercury 3 — API-only, no weights planned. M3.1 Flash weights are promised for October 23, 2026, with the license still undisclosed.

Do they work with my existing OpenAI SDK code?

Mercury 3's API is OpenAI-compatible; M3.1 Flash supports OpenAI- and Anthropic-compatible access. Confirm exact endpoint parameters against each vendor's docs before production use.

Should I trust the benchmark tables?

Treat them as vendor claims: both companies chose the tests and the settings. Independent scores — Artificial Analysis measured M3.1 Flash at 120 tokens/sec and an Intelligence Index of 42–43 — tell a more cautious story.

Key takeaways

• Two speed-first models launched October 19, 2026: Inception's Mercury 3 (diffusion LLM) and MiniMax's M3.1 Flash (MoE agent model).
• Mercury 3 claims 1,068 tokens/sec at $0.10/$0.35 launch pricing; standard rates are $0.20/$0.70 once the promo ends — and the promo has no announced end date.
• M3.1 Flash charges flat $0.15/$0.60/$0.015 with no long-context surcharge, and promises open weights on October 23.
• On vendor-reported numbers, Mercury 3 leads on raw speed while M3.1 Flash leads on agentic coding benchmarks (73.7% on DeepSWE v1.1) — but every benchmark here was run by the vendor selling the model.
• If you build latency-sensitive agents, trial both on your own workload: throughput claims rarely survive contact with a real pipeline.

The October 19 double-launch says something bigger than either model: the small-model tier is now a speed-and-price war, not a quality war. Prices, promos, and weights all move fast here — so measure on your own workload, budget at list price, and read the fine print before the headline. Our roundup of the 10 best new AI tools to try in 2026 is the broader starting point. (Published October 20, 2026.)