Inception launches Mercury 3 with a 1-million-token context window

• Mercury 3 pairs a 1M-token context window with reported generation above 1,000 tokens/s, according to OpenRouter's launch announcement.
• Tool calling and structured outputs are available through AI Gateway, per Vercel's announcement.
• Web research reaches Parallel's price-quality frontier, according to its first diffusion-model evaluation.

Inception scaled Mercury 3 to more than 40 times Mercury 1's parameter count while retaining similar throughput. OpenRouter lists a 131,072-token completion limit beneath that headline 1M context window.

Mercury 3 generates and refines tokens in parallel instead of decoding strictly one token at a time. Inception reports 1,068 output tokens/s on NVIDIA GPUs; OpenRouter's live performance panel shows a 244-token/s median.

The most compelling deployment pattern is a fast worker beneath a smarter orchestrator, the pairing tri_dao said he has wanted.

Inception compares Mercury 3 with the speed-optimized GPT-6 Luna and Claude Haiku 5.5 tier, as TheRundownAI's launch summary notes. All three models use medium reasoning in the comparison.

Mercury 3 is the first diffusion LLM Parallel has tested with its search tool. The company places it on its web-research price-quality frontier.

Parallel reports these results in its October 19 leaderboard:

• Mercury 3: 63.6 Search Intelligence score; $43.40 per 1,000 tasks.
• Claude Haiku 5.5: 62.9; $24.30 per 1,000 tasks.

Adding search raises Mercury 3's score from 28.1 to 63.6, a 35.5-point gain.

In a one-prompt comparison at medium reasoning, daniel_mac8 timed Mercury 3 through OpenRouter at 4.3 seconds versus Haiku 5.5 through claude.ai at 8.0 seconds.

The same prompt through two services: Mercury 3 finishes in 4.3 seconds, Haiku 5.5 in 8.0 seconds.

Inception also published customer measurements:

• TinyFish AI reported 50% success across 50 live Mind2Web navigation tasks, ahead of GPT-6.1-Sol, with token volume and agent-step count matching a frontier model in its production harness.
• Octen AI reported 0.9-second median generation and 2.2-second median end-to-end response time in its Sense API tests.

Context expands from 260K tokens in Mercury 2.5 to 1M in Inception's specifications.

• API availability: Inception, OpenRouter, Baseten, and Vercel AI Gateway.
• Model slug: inception/mercury-3 on OpenRouter and AI Gateway. Vercel's integration uses AI_GATEWAY_API_KEY.
• Tools and JSON: tools and tool_choice for function calling; JSON Schema in response_format for structured outputs, per OpenRouter's API listing.
• Venice: anonymous access through its model page.

At Inception's 50% launch discount, pricing is $0.10/M input and $0.35/M output; standard rates are $0.20 and $0.70 respectively. OpenRouter lists cache reads at $0.01/M.