Key Takeaways
- • Inception Labs released Mercury 3 on 19 October 2026, a diffusion-based language model (dLLM) that generates tokens in parallel, citing 1,068 tokens per second on NVIDIA GPUs, a 1 million token context window, and $0.10 per million input tokens during a launch discount.
- • The launch announcement contains no benchmark scores; its quality claims are comparative, stating Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5, without naming the benchmarks, margins, or per-task results.
- • Inception's only structured quality result is an internal agent test: Mercury 3 completed 50% of 50 Mind2Web navigation tasks, which the company says is well above GPT-6.1-Sol, using its own harness.
- • Independent Artificial Analysis measurements exist only for predecessors (about 706 tokens per second for Mercury 2.5), and as of 20 October 2026 no Mercury 3 page was listed, so the speed claim is unverified.
- • List pricing is $0.20 per million input and $0.70 per million output, with the launch discount having no stated end date; the article recommends planning around list prices.
Summary
Inception Labs released Mercury 3 on 19 October 2026, positioning it as the most capable diffusion-based language model it has released to date. The company highlights three headline figures: throughput of 1,068 tokens per second on NVIDIA GPUs, a context window of 1 million tokens (up from 260K on Mercury 2.5), and a launch price of $0.10 per million input tokens. Maximum output is 131,072 tokens. According to the announcement, the model is more than 40 times larger than Mercury 1 by total parameters, though Inception has not published an absolute parameter count, and measured intelligence across its benchmark suite has roughly tripled from Mercury 1 (mini) to Mercury 3.
The quality claims are comparative rather than quantitative. Inception states Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5, both speed-optimized models, and describes it as matching or exceeding speed-optimized frontier models rather than frontier reasoning models. The announcement does not provide margins, benchmark names, or per-task results. The only structured quality data point is an internal agent evaluation in Inception's own web-agent harness: Mercury 3 completed 50% of 50 Mind2Web navigation tasks, which the company says is far above GPT-6.1-Sol at equal token volume and step count. The sample is small and the harness is proprietary, so the result is a signal rather than a verifiable benchmark.
Standard list pricing is $0.20 per million input tokens and $0.70 per million output tokens, with cached input at $0.02 per million, halved to $0.01 during the launch period. The launch rates of $0.10 and $0.35 have no stated end date. Reasoning effort has four settings (instant, low, medium, high), and Mercury 3 defaults to low when the parameter is omitted, a factor that affects speed and benchmark comparability. Availability is through Inception's own API, which offers 100 million free tokens to new accounts, and several third-party inference platforms. Mercury Voice and Mercury Decide were also introduced as separate endpoints.
Independent verification remains limited. Artificial Analysis previously measured Mercury 2.5 at about 706 tokens per second and Mercury 2 in the 680s, both below vendor-reported figures. As of 20 October 2026, Artificial Analysis had no Mercury 3 listing. The article characterizes the launch as suited to latency-sensitive production workloads with recoverable errors, and notes that the model is not available through OrcaRouter's catalog at the time of writing.
The quality claims are comparative rather than quantitative. Inception states Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5, both speed-optimized models, and describes it as matching or exceeding speed-optimized frontier models rather than frontier reasoning models. The announcement does not provide margins, benchmark names, or per-task results. The only structured quality data point is an internal agent evaluation in Inception's own web-agent harness: Mercury 3 completed 50% of 50 Mind2Web navigation tasks, which the company says is far above GPT-6.1-Sol at equal token volume and step count. The sample is small and the harness is proprietary, so the result is a signal rather than a verifiable benchmark.
Standard list pricing is $0.20 per million input tokens and $0.70 per million output tokens, with cached input at $0.02 per million, halved to $0.01 during the launch period. The launch rates of $0.10 and $0.35 have no stated end date. Reasoning effort has four settings (instant, low, medium, high), and Mercury 3 defaults to low when the parameter is omitted, a factor that affects speed and benchmark comparability. Availability is through Inception's own API, which offers 100 million free tokens to new accounts, and several third-party inference platforms. Mercury Voice and Mercury Decide were also introduced as separate endpoints.
Independent verification remains limited. Artificial Analysis previously measured Mercury 2.5 at about 706 tokens per second and Mercury 2 in the 680s, both below vendor-reported figures. As of 20 October 2026, Artificial Analysis had no Mercury 3 listing. The article characterizes the launch as suited to latency-sensitive production workloads with recoverable errors, and notes that the model is not available through OrcaRouter's catalog at the time of writing.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.