Mercury 3: What Changed, What It Costs, Should You Switch?
Answer first
Inception Mercury 3, released October 19, 2026, is a closed diffusion LLM that drafts a whole answer and refines it in parallel instead of writing token by token. Inception claims 1,068 tokens per second and GPT-6 Luna / Claude Haiku 5.5-class quality at $0.10/$0.35 launch pricing. Independent OpenRouter medians are far lower, so treat the speed as promising, not proven.
What changed vs Mercury 2.5
- A scale jump, not a new idea: Mercury 3 is more than 40× the total parameters of Mercury 1 (June 2025) at comparable speed, per Inception. The exact parameter count is not disclosed.
- Context window 260K → 1M tokens; max output 131,072 tokens (Inception model docs).
- Reasoning effort levels: instant, low, medium, high — tunable per request.
- Quality claim: on Inception's own benchmark table (all runs at medium effort), Mercury 3 beats GPT-6 Luna on 10 of 12 benchmarks and Claude Haiku 5.5 on 10 of 12.
- Migration clock: Mercury 2 and Mercury 2.5 are closing to new users, and existing non-enterprise access ends November 15, 2026 (launch coverage).
- Still text-only: no image or audio input; tool calling and structured outputs are supported.
The numbers, with trust labels
| Benchmark | Mercury 3 | GPT-6 Luna | Claude Haiku 5.5 | Source & trust |
|---|---|---|---|---|
| GPQA Diamond | 89.23 | 86.36 | 84.51 | Inception launch table — vendor-reported, medium effort |
| SWE-Bench Atlas | 51.08 | 33.33 | 32.26 | Inception launch table — vendor-reported |
| DeepSWE v1 | 46.31 | 37.76 | 51.62 | Inception launch table — vendor-reported; Mercury trails Haiku here |
| MRCR v2 (long-context recall) | 75.50 | 62.72 | 78.40 | Inception launch table — vendor-reported; Mercury trails Haiku here |
| Output speed | 1,068 tok/s on NVIDIA GPUs | — | — | Inception measurement — vendor-reported |
| Output speed, real traffic | median 160–254 tok/s (Oct 20, 2026); one-week average median 203 tok/s; median latency ~1s | — | — | OpenRouter live throughput, reported by eesel AI — independent platform measurement |
| Artificial Analysis / LMArena score | none yet | — | — | No AA page exists for Mercury 3 yet; Inception's co-founder said third-party benchmarking is still under way (Hacker News launch thread) |
How to read this table: every quality number above is Inception's own run, at one effort setting (medium). There is no independent quality score to set against them yet. The one independent measurement that does exist — OpenRouter's live throughput — lands at roughly a fifth of the headline speed, which is still fast, and is the number to plan around. Inception's co-founder framed the model the same way on Hacker News: Haiku- or Luna-level in daily use, used successfully for subagents, not yet a replacement for a frontier primary driver.
What it costs (per 1M tokens, checked Oct 23, 2026)
| Route | Input | Output | Cached input | Note |
|---|---|---|---|---|
| Launch price (50% off list) | $0.10 | $0.35 | $0.01 | Same rate on Inception API and OpenRouter |
| List price | $0.20 | $0.70 | $0.02 | Applies when the launch discount ends — Inception has not said when |
| Free allowance | — | — | — | 100 million tokens on new API keys (Inception models page) |
Runs on what?
- API only. Closed weights — nothing to download, self-host, or fine-tune. Access via the Inception API (OpenAI-compatible, model ID
mercury-3), OpenRouter, Baseten, and Vercel AI Gateway. - Text in, text out. No image or audio input, so it cannot sit in a multimodal pipeline slot.
- Speed depends on your traffic shape. Discussion in the launch thread (Hacker News) notes diffusion's latency edge is strongest at low concurrency; Inception says it has scaled well with batch size. Either way: measure on your own workload, at your own effort setting.
Should you switch? By situation
- You run agent loops or subagents on Luna/Haiku-class models → Pilot it. Sequential calls are where output speed compounds; A/B it on your real tasks and measure throughput yourself.
- You are on Mercury 2 or 2.5 → Plan the move now. The old models close to new users and non-enterprise access ends November 15, 2026.
- You need image input, self-hosting, or fine-tuning → Not this model. Text-only and closed.
- You need proven frontier quality → Wait. No independent benchmark has scored it yet, and Inception itself positions it at Haiku/Luna level for daily work.
FAQ
What is a diffusion LLM?
Most LLMs are autoregressive: they predict one token, append it, and repeat, so a 2,000-token answer takes 2,000 sequential steps. A diffusion LLM starts from a rough draft of the whole answer and refines all of it in parallel over a few passes — less typewriter, more editor revising a full draft at once. That is where Mercury 3's speed claim comes from.
Is Mercury 3 really 1,068 tokens per second?
That is Inception's measurement on its own NVIDIA hardware. OpenRouter, which routes all Mercury 3 traffic to Inception, measured medians of 160–254 tokens per second on October 20, 2026, with a one-week average median of 203 and median latency around one second. Still fast versus the ~130 tok/s Inception lists for GPT-6 Luna — but plan around the measured number.
Is Mercury 3 open source?
No. It is a closed, proprietary model: no weights are published, so it cannot be self-hosted or fine-tuned. If open weights matter to you, this is a rental-only option.
When does the 50% launch discount end?
Inception has not announced an end date as of October 23, 2026. Budget at the list price ($0.20 input / $0.70 output per 1M tokens) and treat the discount as temporary.
Can Mercury 3 replace a frontier model like Claude Opus?
Not on current evidence. Inception's own table compares it against speed-tier models (GPT-6 Luna, Claude Haiku 5.5), and its co-founder describes it as Haiku- or Luna-level in daily use, successful as a subagent engine rather than the primary driver of a harness.
What we could not verify
- Parameter count: Inception has not disclosed it (only "more than 40× Mercury 1")
- Any independent quality benchmark: Artificial Analysis and LMArena had not scored Mercury 3 as of Oct 23, 2026
- The end date of the 50% launch discount
- Release-date discrepancy: OpenRouter's listing shows Oct 16, 2026 while the launch was Oct 19, 2026
Sources: Inception launch post and model docs (figures as compiled in eesel AI's Oct 20, 2026 explainer, checked Oct 23, 2026); OpenRouter live throughput data for Oct 20, 2026 (via eesel AI); Unite.AI launch coverage, Oct 19–20, 2026; Venice AI model page (access and pricing). Vendor-reported figures are labelled as such wherever they appear on this card.