Mercury 3 vs Mercury 2: Which Diffusion Model Fits Your Stack?

Mercury 3 is Inception's newest model and Mercury 2 is its February 2026 predecessor, so the question is whether to move a workload from Mercury 2 to Mercury 3, or to start on Mercury 3.

The practical gaps are context (1M tokens vs 128K), maximum output (131,072 vs 50,000 tokens), regular price ($0.20/$0.70 vs $0.25/$0.75 per million tokens) and the reasoning default (low vs medium).

Mercury 3 vs. Mercury 2: Side-by-Side

Are you one of these vendors? Update your listing

Mercury 3 or Mercury 2: The Short Answer

Choose Mercury 3 for any new deployment, long-context workload, or pipeline requiring extended generation outputs. Inception lists Mercury 3 with a 1M-token context window, 131,072 maximum output tokens, and lower list pricing than its predecessor, calling it its most capable diffusion-based language model to date (Diffusion at Scale: Introducing Mercury 3).

Keep Mercury 2 only if your team maintains an existing, tested workload that fits comfortably within 128K tokens and you do not wish to re-evaluate prompt behaviors or reasoning defaults today. The deciding question is whether your production application requires expanded context, longer generated answers, or lower API token bills, and whether you have allocated engineering time to re-run quality evaluations against the new model identifier. More detail is in Mercury 3 explained.

• Pick Mercury 3 if your prompts exceed 128K tokens or require up to 131,072 output tokens.
• Pick Mercury 3 if you want lower base API rates ($0.20 input and $0.70 output per million tokens).
• Keep Mercury 2 if your application relies on locked prompts tuned specifically to Mercury 2's default medium reasoning effort.
• Consider Mercury 2.5 as an intermediate step if you need 260K tokens of context without altering default reasoning behavior.

What Changed From Mercury 2 to Mercury 3

Inception released three models in this line during 2026. Inception introduced Mercury 2 on February 24, 2026, as a speed-focused reasoning model, with Inception reporting 1,009 tokens per second on NVIDIA Blackwell GPUs (Introducing Mercury 2). On September 8, 2026, Inception released Mercury 2.5, claiming a 40 percent increase in intelligence over Mercury 2 and expanding context capacity to 260K tokens (Introducing Mercury 2.5). On October 19, 2026, Inception announced Mercury 3, scaling parameter capacity more than 40 times relative to Mercury 1.

Inception says Mercury's models generate tokens in parallel through diffusion instead of left to right (Inception blog). Both models list tool calling and structured outputs, and both are served through the OpenAI-compatible chat completions endpoint at v1/chat/completions using the base URL https://api.inceptionlabs.ai/v1.

If you omit the reasoning_effort parameter in your request payload, Mercury 3 defaults to low, whereas Mercury 2 defaults to medium.

• Parameter scale: Inception claims Mercury 3 is over 40 times larger in parameters than Mercury 1; no parameter count is published for Mercury 3.
• Context window: expanded from 128K tokens on Mercury 2 to 1M tokens on Mercury 3.
• Output ceiling: expanded from 50,000 tokens on Mercury 2 to 131,072 tokens on Mercury 3.
• Reasoning default: changed from medium on Mercury 2 to low on Mercury 3.
• Unchanged features: OpenAI-compatible API, text-only input, tool calling and structured outputs, per Inception's docs and posts.

Price Comparison and Cost Calculation

Mercury 3 provides lower list pricing across both prompt inputs and generation outputs compared to Mercury 2. As of October 2026, Inception lists Mercury 2 at $0.25 per 1,000,000 input tokens, $0.025 per 1,000,000 cached input tokens, and $0.75 per 1,000,000 output tokens.

As of October 2026, Inception also lists a launch promotion for Mercury 3 on its documentation home page (Inception Documentation Home). This promotion cuts rates by 50 percent to $0.10 per million input tokens, $0.01 per million cached input tokens, and $0.35 per million output tokens. Inception publishes no expiration date for this launch discount.

Our arithmetic for a workload of 1,000,000 input tokens and 1,000,000 output tokens: On Mercury 2, this volume totals $1.00 ($0.25 input plus $0.75 output). On Mercury 3 at standard rates, the same volume totals $0.90 ($0.20 input plus $0.70 output). Under the 50 percent launch discount, Mercury 3 costs $0.45 ($0.10 input plus $0.35 output). Total cost also depends on token efficiency and reasoning token generation, which vary by task and reasoning effort tier.

For the full rate card, see Mercury 3 pricing and Mercury 2 API pricing.

• Mercury 2 regular cost: $0.25 input / $0.025 cached / $0.75 output per 1M tokens.
• Mercury 3 regular cost: $0.20 input / $0.02 cached / $0.70 output per 1M tokens.
• Mercury 3 launch cost: $0.10 input / $0.01 cached / $0.35 output per 1M tokens.
• Standard 1M in / 1M out sample workload: $1.00 on Mercury 2 vs $0.90 on Mercury 3 regular ($0.45 launch).
• Cached inputs on both models receive a 90 percent discount relative to standard input rates.

→ Compare per-token pricing for every AI model

Speed and Quality Claims From Inception's Posts

Speed and throughput claims for both models come directly from Inception's release announcements rather than identical head-to-head laboratory tests. In its February 2026 launch post, Inception reported Mercury 2 generating 1,009 tokens per second on NVIDIA Blackwell GPUs, describing it as competitive with leading speed-optimized models (Introducing Mercury 2). For Mercury 3, Inception reported 1,068 tokens per second on unspecified NVIDIA GPUs in October 2026 (Diffusion at Scale: Introducing Mercury 3).

Because these figures come from separate Inception posts that do not state matching hardware, prompts or settings, they do not establish that one model is definitively faster than the other in production. Inception publishes no direct head-to-head latency benchmark between Mercury 3 and Mercury 2. For Mercury 2.5, Inception published a figure of 1,107 tokens per second on widely available NVIDIA GPUs, which is numerically higher than Mercury 3's reported 1,068 tokens per second, but hardware configurations were not stated as matched.

Regarding quality, Inception states that Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks while running four to eight times faster, though no benchmark score tables or technical reports are published. Inception ran a web-agent test where Mercury 3 completed 50 percent of 50 live Mind2Web navigation tasks in a TinyFish harness, outperforming GPT-6.1 Sol while matching frontier step counts.

• Mercury 2 speed: reported at 1,009 tokens/sec on NVIDIA Blackwell GPUs in February 2026.
• Mercury 3 speed: reported at 1,068 tokens/sec on NVIDIA GPUs in October 2026.
• Hardware parity: benchmark hardware environments and prompt sets are not documented as like-for-like.
• Mercury 3 benchmark claims: stated to match or exceed GPT-6 Luna and Claude Haiku 5.5 without published score tables.
• Agent navigation: completed 50% of 50 Mind2Web tasks in an Inception internal evaluation harness.

Mercury 2.5 as the Intermediate Option

Inception launched Mercury 2.5 on September 8, 2026 (Introducing Mercury 2.5). Inception positioned Mercury 2.5 as delivering a 40 percent increase in intelligence over Mercury 2, framing it against cost-optimized frontier models such as GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5.

Mercury 2.5 sits between the two in context and output length, and its $0.75 output price matches Mercury 2 while its $0.20 input price matches Mercury 3. It expands context capacity from Mercury 2's 128K tokens to 260K tokens and increases maximum generation from 50,000 tokens to 65,536 tokens.

Mercury 2.5 retains Mercury 2's default medium setting for reasoning_effort. Mercury 2.5 keeps the medium default, so a team that wants more context than 128K and does not want to change the reasoning default can test it first.

• Release timing: September 8, 2026, six weeks prior to Mercury 3.
• Context and output limits: 260K input tokens and 65,536 maximum output tokens.
• Pricing structure: $0.20 per 1M input tokens and $0.75 per 1M output tokens.
• Reasoning default: retains medium as the default reasoning effort, matching Mercury 2.
• Serving speed: Inception reported 1,107 tokens per second on widely available NVIDIA GPUs at launch.

Moving Workloads and Re-Testing Requirements

Upgrading from Mercury 2 to Mercury 3 means setting the model ID to mercury-3. Inception says Mercury 3 is OpenAI API compatible and "drops into your existing stack with no rewrites" (Diffusion at Scale: Introducing Mercury 3).

The behavioral change is the reasoning_effort parameter. Mercury 2 defaults to medium, whereas Mercury 3 defaults to low when this parameter is left unspecified. If your application relies on implicit defaults, moving to Mercury 3 applies low reasoning effort instead of medium. To perform an accurate comparative evaluation, set reasoning_effort to medium explicitly in your Mercury 3 requests.

Engineering teams must also evaluate context thresholds and rate tier boundaries. High-volume applications hitting 429 status codes should implement exponential backoff algorithms.

• Model ID: set to mercury-3.
• Reasoning effort alignment: set reasoning_effort to medium explicitly if replicating Mercury 2 behavior.
• API rate limits: monitor consumption against Inception's Pay As You Go cap of 3,000,000 input tokens per minute.
• Endpoint: Mercury 3 uses the same v1/chat/completions endpoint.
• Temperature adjustments: note that Mercury 3 supports temperature adjustments between 0.5 and 1.0.

Who Should Stay and Who Should Move

The deciding question is whether the workload is already tested on Mercury 2 and whether it needs more context or longer output.

• Stay on Mercury 2 if your production prompts are fragile, tightly calibrated to Mercury 2's reasoning patterns, and fit within 128K tokens.
• Stay on Mercury 2 if your engineering team lacks bandwidth to run regression evaluations against new model outputs.
• Move to Mercury 3 if you wish to lower input and output API bills under regular rates ($0.20/$0.70) or use the 50 percent launch discount ($0.10/$0.35).

The Verdict

For engineering teams starting fresh projects or building applications that handle large context windows, Mercury 3 is the recommended choice. Its 1M-token context capacity, 131,072 max output ceiling, and reduced baseline pricing ($0.20 input and $0.70 output per million tokens) give it higher limits and lower rates than Mercury 2. As of October 2026, the 50 percent launch discount, which has no published end date, further widens the cost advantage.

For mature systems running in production on Mercury 2, the decision requires caution. Inception calls Mercury 3 its most capable model, but Mercury 3 changes the default reasoning effort from medium to low when the parameter is omitted. Teams with fragile prompts should test Mercury 3 in a staging environment by setting reasoning effort explicitly to medium before cutting over live production traffic.

The pick changes if Inception publishes a Mercury 3 vs Mercury 2 head-to-head or a technical report that shows a regression on your kind of task, or if it changes the regular Mercury 3 price. Until then, run your own evals on your own data.