Mercury 3: Inception's diffusion LLM claims to beat GPT-6 Luna and Claude Haiku 5.5 at over 1,000 tokens/s

Mercury 3: Inception's diffusion LLM claims to beat GPT-6 Luna and Claude Haiku 5.5 at over 1,000 tokens/s

🔎 A diffusion model that takes on the giants head-on

On October 19, 2026, Inception Labs, the startup founded by Stefano Ermon, a professor at Stanford, unveiled Mercury 3, a language model that doesn't work like the others. Where GPT, Claude, and Gemini generate text token by token, from left to right, Mercury 3 produces text through diffusion — in the manner of image generators — by iteratively refining entire blocks of text.

The result claimed by the company: 1,068 tokens per second on Nvidia GPUs, a speed 4 to 8 times higher than that of autoregressive models optimized for latency. And this is no second-rate model: Inception claims that Mercury 3 achieves qualitative parity with OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5, while consuming a context of 1 million tokens, compared to 260k for its predecessor Mercury 2.5.

If the numbers hold up, this is a small architectural revolution. But it must be said right away: these benchmarks are self-published. No independent replication is available to date. That is the point we are going to examine carefully.

Key Takeaways

• Mercury 3 is a diffusion LLM: it generates text through parallel block refinement, rather than token by token, allowing it to reach 1,068 tokens/s on Nvidia GPUs — 4 to 8 times faster than optimized autoregressive models.
• Claimed quality on par with GPT-6 Luna and Claude Haiku 5.5, with a 1 million token context window (versus 260k for Mercury 2.5) and a size 40 times larger than Mercury 1.
• Aggressive launch pricing: $0.10/M input tokens and $0.35/M output tokens, positioning Mercury 3 in the budget model range, despite claimed premium performance.
• Warning: all figures come from Inception Labs itself. No independent replication to date. Caution is warranted before making any architecture decisions.

Recommended Tools

To see where Mercury 3 fits in the full landscape, check out our monthly comparison of the best LLMs and our selection of the best LLMs for coding.

How a diffusion LLM works — and why it's faster

Direct answer: a diffusion model generates an entire block of text at once, then refines it iteratively, instead of predicting one token after another. It's this parallelization that explains the speed gains.

Classic LLMs (GPT, Claude, Gemini, Llama) are autoregressive: each token depends on the previous one. It's a sequential chain — it's impossible to generate token 500 before token 499, even with powerful GPUs. Generation latency is therefore structurally bounded by this dependency.

Diffusion reverses the logic. The model starts from a noisy text — a sequence of random tokens — and denoises it in successive steps, correcting the entire block in parallel. In a few refinement passes, the text converges toward a coherent output. Each pass processes dozens or hundreds of tokens simultaneously.

This is the approach that made image generators successful (Stable Diffusion, DALL-E). Transposing it to language was a known challenge: text is discrete, not continuous, and local errors propagate differently. Inception Labs, co-founded by Stefano Ermon, one of the academic pioneers of diffusion, is the team best positioned to solve it.

The historical trade-off of this approach: quality per pass. A diffusion model must "guess" entire blocks before correcting them, which can produce local inconsistencies. Mercury 3's claim is to have closed this gap — precisely what independent benchmarks will have to verify.

The announced figures: 1,068 tokens/s, 1M context, 40x bigger

Direct answer: Inception claims 1,068 tokens/s on Nvidia GPUs, a 1 million token context, and a model 40 times larger than Mercury 1. Here's what to take away from each figure.

Speed: 1,068 tokens/s. For comparison, autoregressive models optimized for speed — GPT-6 Luna, Claude Haiku 5.5, Gemini Flash — typically run between 130 and 250 tokens/s in production on comparable infrastructure. A 4 to 8x factor, if confirmed, is a game changer for use cases where latency is the number one criterion: agents in loops, interactive code generation, real-time voice.

Context: 1 million tokens. Mercury 2.5 capped out at 260k. This jump brings Mercury 3 close to the best on the market on this criterion and makes it usable for analyzing entire corpora, large codebases, or legal files. Worth noting: the Claude 5.5 family recently completed by Anthropic is also betting on 1 million tokens of context with Haiku 5.5 — the long-context race is now open on all fronts.

Size: 40x Mercury 1. A massive leap between two generations, which partly explains the quality gains. Inception is not disclosing (at this stage) the exact parameter count or size in bytes — a regrettable opacity for judging the real performance/infrastructure cost ratio.

Quality: parity with GPT-6 Luna and Haiku 5.5. This is the strongest claim, and the one most in need of verification. Reaching parity with the big labs' "fast" models — not their flagship models — is already ambitious. Note that GPT-6 Luna is the little brother of a family that halved prices upon launch, and that Haiku 5.5 is Anthropic's entry-level model. Mercury 3's target is therefore clear: the speed/price segment, not the heavy reasoning segment.

Price: $0.10/M input, $0.35/M output — a cost offensive

Direct answer: at $0.10 per million input tokens and $0.35 per million output tokens, Mercury 3 positions itself at the low end of the market, while claiming mid-range performance.

This pricing strategy is consistent with Inception's approach: diffusion reduces infrastructure cost per generated token, since parallelism improves GPU utilization. If the model delivers on its quality promises, the quality/price/speed ratio becomes hard to beat.

Let's put this in market perspective. The price war is already underway: OpenAI has positioned GPT-6 Sol and Luna at half the price of the previous generation, DeepSeek continues to slash prices with V4 Pro and Flash, and Chinese models like Alibaba's new Qwen3.6 family are squeezing the market from below. In this context, $0.10/$0.35 is competitive but not revolutionary — it's the combination of speed + price that makes the argument.

The math is simple for an agentic use case: an agent generating 10 million tokens per day would pay about $3.50/day with Mercury 3, versus potentially 4 to 8 times more with an autoregressive model of equivalent quality — not counting the latency gain, which multiplies the number of possible iterations per hour.

What this changes for developers and AI agents

Direct answer: if the claims hold up, Mercury 3 becomes a serious candidate for agentic architectures and applications where latency is a blocker.

AI agents are the most obvious use case. An agent chaining 20 tool calls with reasoning between each pass loses considerable time on sequential generation. At 1,068 tokens/s, the model's reasoning time becomes almost negligible compared to tool execution time. For this workload profile, our guide to the best LLMs for AI agents will probably need updating.

Second use case: code. The IDE completion experience is extremely latency-sensitive — beyond a few hundred milliseconds, suggestion acceptance rates drop. A model 4 to 8 times faster, with a 1M token context to absorb an entire codebase, checks both boxes. For now, Claude Sonnet 4.6 and GPT-5.3 Codex dominate this segment — Mercury 3 could shake things up.

Third use case: consumer applications with streaming. Content generation, voice assistants, chatbots — anywhere the user expects an instant response, decoding speed is a direct product advantage.

On the other hand, for deep reasoning — mathematics, scientific research, complex analysis — heavy reasoning models (GPT-5.5, Gemini 3 Pro Deep Think, Claude Opus 4.7) remain in a different category. Mercury 3 doesn't claim to compete with them, and we shouldn't attribute intentions it doesn't have.

The flip side: self-published benchmarks

Direct answer: all the figures in this launch come from Inception Labs. No independent evaluation confirms them to date. This is the most important point of this article.

This is not an accusation — it's the industry norm. OpenAI, Google, and Anthropic also publish their own benchmarks at launch. But three things deserve particular vigilance here.

First, the "parity with GPT-6 Luna and Haiku 5.5" comparison depends entirely on the benchmarks chosen and the test parameters. A 4x to 8x speed factor can be measured under favorable conditions: aggressive batching, short prompts, cutting-edge hardware not precisely specified. Real-world production speed, with heterogeneous requests, is almost always lower.

Second, diffusion has a known blind spot: tasks that require very long-range coherence — multi-step reasoning, structured code spanning hundreds of lines — have historically been harder for this architecture. Do the parity benchmarks cover these tasks, or general generation tasks? Inception hasn't publicly detailed this yet.

Finally, the precedent is encouraging but limited: previous versions of Mercury had been more or less independently tested and delivered on their speed promises, with quality lagging behind the leaders. Mercury 3 claims to have closed that gap — this is exactly the kind of claim that needs to be replicated before being believed.

Our advice: wait two to three weeks. Community benchmarks (LiveBench, Artificial Analysis, independent evaluations) will arrive quickly. If you have a use case where speed is critical, test Mercury 3 in parallel with your current solution on your own data — that's the only benchmark that really matters.

Where Mercury 3 Stands Against the Competition

Direct answer: Mercury 3 targets the speed/price segment, dominated by GPT-6 Luna, Claude Haiku 5.5, and Gemini Flash — not the frontier reasoning model segment.

The speeds of autoregressive models are market estimates, not official figures — they vary depending on the inference provider, batching, and load.

The table reveals Inception's strategy: not to fight on raw quality against GPT-5.5 or Claude Opus 4.7, but to redefine the speed/quality/price ratio in the segment where latency is the decisive factor. This is a real and growing niche, driven by the rise of agentic architectures.

Also worth watching: the reaction of the major labs. If diffusion proves its viability at scale, expect Google, OpenAI, or the Chinese labs to explore hybrid architectures — that's already the direction academic research has been taking for the past two years.

❌ Common Mistakes

Mistake 1: Taking launch benchmarks at face value

The figures published at launch are measured by the company itself, under conditions it chooses. Solution: wait for independent replications (usually 2-3 weeks) and test on your own use cases before migrating.

Mistake 2: Confusing tokens/s with perceived latency

1,068 tokens/s doesn't mean the first response arrives instantly. Time-to-first-token depends on prefill and initial inference. Solution: measure end-to-end latency on your application, not just throughput.

Mistake 3: Migrating an entire stack for a speed gain

Switching models for speed without validating quality on your real data is the best way to silently degrade your product. Solution: A/B test Mercury 3 on 10% of your traffic before any switchover.

Mistake 4: Forgetting switching costs

Changing models involves adapting prompts, revalidating outputs, and managing behavioral differences. The gain of $0.05/M tokens doesn't always justify this cost. Solution: calculate the total migration cost before deciding.

❓ Frequently Asked Questions

Is Mercury 3 really faster than GPT-6 Luna?

According to Inception Labs, yes: 4 to 8 times faster, with 1,068 tokens/s measured on Nvidia GPUs. But these are self-published figures, not independently replicated to date. The measurement conditions (batching, sequence length, exact hardware) are not fully detailed. Caution is warranted.

Can Mercury 3 replace Claude or GPT for coding?

Not yet demonstrated. The claimed parity concerns speed-optimized models (Haiku 5.5, GPT-6 Luna), not flagship coding models like Claude Sonnet 4.6 or GPT-5.3 Codex. For serious coding, check out our comparison of the best LLMs for coding and wait for independent tests on Mercury 3.

What is a diffusion LLM, in simple terms?

A model that generates text by starting from a noisy draft that it corrects through successive passes, like an image generator. Unlike traditional LLMs that write word by word, it refines entire blocks in parallel — hence the speed, but also challenges with long-range coherence.

Is the $0.10/M tokens price final?

It's a launch price (October 2026, check inceptionlabs.ai). Launch prices in this sector rise or fall quickly depending on demand and competition. Don't build a long-term cost strategy on a launch price.

Should you adopt Mercury 3 right now?

Only if latency is your absolute constraint and you can test in real-world conditions. For most use cases, wait for the independent benchmarks — two to three weeks of patience beats a premature migration.

✅ Conclusion

Mercury 3 is the most architecturally interesting launch of the quarter: diffusion finally scales to industrial levels, with figures — 1,068 tokens/s, 1M context, $0.10/M input — that, if they hold up, redefine the speed/price segment against GPT-6 Luna and Claude Haiku 5.5. But everything rests on self-published benchmarks: test it, compare, and follow our monthly comparison of the best LLMs to see whether Mercury 3 delivers on its promises once it passes the independence filter.