Inception Labs claims Mercury 3 beats Haiku and GPT-6 Luna at 1,068 tokens a second

Inception Labs released Mercury 3 on October 19, claiming the diffusion-based model beats Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6 Luna on IFBench, GPQA Diamond and SWE-Bench Atlas while running at 1,068 tokens per second on Nvidia GPUs. OpenRouter confirmed Mercury 3 went live on its platform the same day, citing "1,000+ tok/s with 1M context, 4x more than Mercury 2.5," and Inception priced the model at $0.10 per million input tokens and $0.35 per million output tokens, a launch discount the company says is 50% off its standard rate. All benchmark and throughput figures are vendor-reported and have not been replicated by an independent evaluator, the same status Mercury 2.5's claims carried at its September launch.

Inception Labs says its newest model generates text four to eight times faster than rivals while matching their quality, a combination the industry has long treated as impossible to pull off at the same time. Inception Labs released Mercury 3 on October 19, claiming it beats both Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6 Luna on three benchmarks, IFBench, GPQA Diamond, and SWE-Bench Atlas, while running at 1,068 tokens per second on Nvidia GPUs. OpenRouter confirmed the model went live on its platform the same day, posting that Mercury 3 generates "1,000+ tok/s with 1M context, 4x more than Mercury 2.5." The company priced it at $0.10 per million input tokens and $0.35 per million output tokens at launch, a discount the firm says is 50% off its standard rate. That price and speed combination matters because Inception isn't just another startup chasing GPT-6 on quality. It's the company that built the entire commercial case for diffusion-based language models in the first place, and Mercury 3 is its biggest bet yet that the approach scales. Every major foundation model from OpenAI, Anthropic, and Google generates text one token at a time, each word conditioned on everything before it. That's the autoregressive approach, and it's reliable but slow, because the model can't start word six until it's finished word five. Inception, founded by Stanford's Stefano Ermon along with Volodymyr Kuleshov and Aditya Grover, bet on a different method borrowed from image generation: diffusion. Instead of writing left to right, a diffusion model starts with a noisy draft of the entire response and refines all of it in parallel across a handful of steps, the same coarse-to-fine process that turns static into a photo in tools like Midjourney. The company's first Mercury model, launched in 2025, already hit over 1,000 tokens per second on Nvidia H100s while matching models like Gemini 2.0 Flash-Lite and GPT-4o-mini on coding benchmarks. Mercury 2 followed with reasoning capability and a quality level the company says sits alongside Claude Haiku 4.5, at roughly 89 tokens per second, and GPT-5 Mini, at around 71. Mercury 3 is the next jump: Inception says it's more than 40 times larger than the original Mercury while holding the same throughput, because scaling the parallel refinement process doesn't carry the same speed penalty that scaling an autoregressive model does. Goldman Sachs Upgrade Sends CoreWeave and the AI Trade Bouncing Back https://startupfortune.com/goldman-sachs-upgrade-sends-coreweave-and-the-ai-trade-bouncing-back/ Goldman's Gabriela Borges set a $106 price target on CoreWeave, implying 73% upside, even as she cut it from $139, while the Philadelphia Semiconductor Index snapped an 18-year losing streak. - Goldman Sachs upgrades CoreWeave stock to buy rating https://startupfortune.com/goldman-sachs-upgrade-sends-coreweave-and-the-ai-trade-bouncing-back/ - CoreWeave stock price target and analyst upgrade impact https://startupfortune.com/goldman-sachs-upgrade-sends-coreweave-and-the-ai-trade-bouncing-back/ The context window grew just as sharply. Mercury 2.5 topped out around 260,000 tokens. Mercury 3 handles a full million, putting it in the same range as Google's Gemini line and well past what most speed-tier models offer. The catch nobody's confirmed yet Here's the thing: every one of those numbers comes from Inception itself. The company's own benchmark comparisons against GPT-6 Luna and Claude Haiku 5.5 haven't been replicated by an independent evaluator, and that pattern isn't new. Mercury 2.5's throughput claims carried the same label when it launched in September, vendor-reported and unverified by any third party at the time. A model can post whatever score it wants on its own test harness. The real test is what happens when developers run their own workloads against it, and Inception is betting that happens fast, because it's already shipping the model everywhere developers look: OpenRouter, Baseten, Vercel, and its own API, all OpenAI-compatible so switching providers takes a config change, not a rewrite. Inception says early-access partners are already running Mercury 3 in browser-use agents, interactive education tools, real-time search agents, and customer support systems, exactly the latency-sensitive categories where a slow model breaks the user experience regardless of how smart it is. That's the real argument for diffusion over scale: a coding assistant that takes four seconds to finish a thought loses developers to one that takes half a second, even if the slower model is marginally sharper. If Mercury 3's numbers hold up under outside testing, it reframes a pricing argument the big labs haven't had to fight yet. OpenAI and Anthropic have spent two years competing on reasoning depth and context length. Inception is making the case that none of that matters if the response takes too long to land, and at a third of the going rate, it's daring the market to check. Also read: Goldman Sachs Upgrade Sends CoreWeave and the AI Trade Bouncing Back https://startupfortune.com/goldman-sachs-upgrade-sends-coreweave-and-the-ai-trade-bouncing-back/ • Sixteen Hours to Number One on Bing Copilot: An Irish Agency Tests AI Visibility on Itself https://startupfortune.com/sixteen-hours-to-number-one-on-bing-copilot-an-irish-agency-tests-ai-visibility-on-itself/ • US Bars DeepSeek From Civilian Federal Agencies and Government Contracts https://startupfortune.com/us-bars-deepseek-from-civilian-federal-agencies-and-government-contracts/ This article is posted in AI News https://startupfortune.com/category/ai/ , check it out for more related stories. Join the discussion Open in the community → https://startupfortune.com/community/ Almost there. Sign in and your reply posts straight away.