Tool Suggestions
Mercury 3: The Diffusion LLM That Writes at 1,068 Tokens per Second
Every AI agent waits on its model. An agent that plans, calls a tool, reads the result and decides again may call the model ten times for one task, and each call costs seconds. On October 19, 2026, Inception released Mercury 3, a language model that attacks that wait from a different direction: instead of writing text one token at a time like GPT-6 Luna or Claude Haiku 5.5, it drafts a whole block of text in parallel and refines it over a few passes — the diffusion approach long used by image generators, applied to language. Inception claims 1,068 output tokens per second, a 1-million-token context window and quality on par with the speed-tier frontier models. This article sets out what shipped, what the pricing is, and where the claims still rest on the company's own measurements.
What Mercury 3 is
Mercury 3 is the fourth release in Inception's diffusion LLM line since Mercury 1 in June 2025, and the company says it is more than 40 times larger than that first model, though the exact parameter count is not disclosed. The context window grows to 1 million tokens, up from 260,000 in Mercury 2.5. It accepts text only, supports tool calling and structured outputs, and is served through an OpenAI-compatible API on Inception's own platform as well as OpenRouter, Baseten and the Vercel AI Gateway. The weights are not released. Mercury 2 and 2.5 are closed to new users, and existing customers keep access only until November 15, 2026 — Mercury 3 is now the product. Inception's backers include Menlo Ventures, Microsoft's M12, Andrew Ng and Andrej Karpathy, and co-founder Stefano Ermon's launch claim was direct: the limits to scaling diffusion models are "compute and data, not fundamental algorithmic barriers."
Mercury 3 pricing: the cheapest output in its class, for now
The list price is $0.20 per million input tokens and $0.70 per million output tokens, with cached input at $0.02. A launch promotion halves those rates to $0.10 input and $0.35 output, with no announced end date, and new API keys come with 100 million free tokens. Against its named rivals — GPT-6 Luna and Claude Haiku 5.5, both at $0.10 input and $0.50 output — Mercury 3 matches on input and undercuts output by 30% while the discount lasts. The honest reading is that the discount is doing much of the work: at list price, Mercury 3's output is the most expensive of the three. Teams evaluating it should model both prices.
The speed claim vs independent measurement
The headline figure — 1,068 tokens per second — comes from Inception's own testing on NVIDIA GPUs; the company has not said which GPU, and its own pages variously describe the advantage as 5x and as 4 to 8x over Luna and Haiku 5.5. The first independent datapoint arrived quickly: OpenRouter, which routes Mercury 3 traffic to Inception, measured a median of 160 to 254 tokens per second on October 20, with a one-week average of 203. That is far below the headline — and still roughly at the level of Claude Haiku 5.5's vendor figure of about 240 tokens per second, delivered by a model whose selling point is its price class, not its architecture. Early adopter reports point the same way: Octen AI measured a median generation time of 0.9 seconds, and TinyFish reported Mercury 3 completing 50% of 50 live web-agent tasks, ahead of GPT-6.1 Sol in its test. The fair conclusion is that Mercury 3 is genuinely fast in production, and that the 1,068 figure describes conditions buyers should not plan around.
Benchmarks: what the comparison table does and does not show
On Inception's own benchmark table, run at medium effort, Mercury 3 scores higher than GPT-6 Luna and Claude Haiku 5.5 on 10 of 12 tests. The exceptions are informative: it trails Luna on AA-Omniscience and AA-LCR, and trails Haiku 5.5 on MRCR v2 and DeepSWE — long-context retrieval and deep software engineering, two workloads agent builders care about. No independent leaderboard had scored Mercury 3 at the time of writing, and Inception has not published the full benchmark list behind its "on par or better" claim. Treat the table as the company's case, not a verdict.
Who should try Mercury 3
The fit is workloads where latency compounds: multi-step agents that call the model in a loop, voice agents where a five-second pause reads as broken, and real-time interfaces such as tutoring and search. With 100 million free tokens for new keys and an OpenAI-compatible API, testing it against an existing agent workload costs an afternoon rather than a migration. It is a poorer fit where long-context retrieval or maximum coding depth is the bottleneck, and teams should note the text-only input — there is no image understanding. Anyone whose workload is cost-bound rather than speed-bound should also price the comparison in our Claude Haiku 5.5 analysis before deciding.
The takeaway
Mercury 3 is the most credible challenge yet to the assumption that language models must write one token at a time. The architecture is real, the production speed is genuinely high, and the launch pricing is aggressive. The headline numbers — 1,068 tokens per second, "better than Luna and Haiku" — are vendor-reported and already look softer under independent measurement. For agent builders, that still leaves a model worth an afternoon's evaluation, which is more than most launches earn.