Mercury 3 vs Claude Opus 5.5: A 57-Times Cheaper Output Token Is Not the Same as a Cheaper Answer

Mercury 3 costs $0.35 per million output tokens this week. Claude Opus 5.5 costs $20.00. Multiply a support queue by those two numbers and the gap looks like a mistake in one of the rate cards — 57 times, at the launch price, or 28.6 times once Inception's 50% discount ends and Mercury 3 reverts to $0.70. The question that decides the comparison is not which number is smaller. It is how many Mercury 3 calls you have to make to get an answer that Opus 5.5 would have got right the first time, because every retry, every escalation and every human review is billed at the rate of the person or the second call, not at the rate of the token. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

That framing matters because these two models are not competitors in the ordinary sense. Claude Opus 5.5, released by Anthropic on September 22, 2026, is a frontier reasoning model at $4 per million input and $20 per million output, built for long-horizon agentic coding and knowledge work, with adaptive thinking, vision and file input, a 1M-token context and a 128K output ceiling. Mercury 3, released by Inception Labs on October 19, 2026, is a diffusion language model built for speed: 1,068 tokens per second on NVIDIA hardware by Inception's own measurement, 1M context, 131K output, text only, and a quality claim that names GPT-6 Luna and Claude Haiku 5.5 — the speed tier — rather than Opus. Comparing them is still worth doing, because the price gap invites exactly the substitution that the positioning does not support.

Where the numbers actually come from

Every Opus 5.5 quality figure in this piece is Artificial Analysis's, measured on its own harness: an Intelligence Index of 57.6, Humanity's Last Exam at 61.4%, SciCode at 66.9%, Terminal-Bench 4.0 at 59.6% and long-context recall at 84.7%. Those are independent, and they are current as of this week.

Every Mercury 3 figure is Inception's, published yesterday, with one exception: the throughput comparison to the previous generation. Artificial Analysis measures Mercury 2.5 at 705.9 tokens per second, and Mercury 2 at roughly 684. Mercury 3 has no Artificial Analysis page yet — the model index still stops at Mercury 2.5 — so its 1,068 tok/s and its "outperforms GPT-6 Luna and Claude Haiku 5.5" claim are vendor-reported, unreproduced, and should be read as such for at least another week.

That asymmetry is the whole comparison in miniature. One side has an independently audited score and a published rate card; the other has a launch post, a discount and a very large speed number.

Spec by spec, with the evidence attached

• Output price — Mercury 3 $0.35 per million at launch, $0.70 at list, against Opus 5.5's $20.00. The ratio is 57x during the promotion and 28.6x after it.

• Input price — Mercury 3 $0.10 at launch, $0.20 at list, against Opus 5.5's $4.00. Same pattern, slightly wider ratio.

• Cached input — Mercury 3 $0.01 at launch, $0.02 at list, against $0.20 for Opus 5.5 cache reads. Opus also charges $5.00 per million to write to the 5-minute cache and $8.00 for the one-hour cache, which Mercury's rate card does not break out.

• Throughput — Mercury 3 at 1,068 tok/s per Inception, unreproduced; Opus 5.5 at 75.6 tok/s output as measured across our own playground traffic over the last seven days. That is a fourteen-fold gap in the direction Inception claims, and this one is at least measured on both sides, albeit by different harnesses.

• Time to first token — Mercury 3 is a diffusion model that generates in parallel, so its first useful output arrives sooner than an autoregressive model's; Inception does not publish a figure for the chat endpoint, and its sub-300ms claim on the company homepage refers to the product line rather than to Mercury 3 specifically. Opus 5.5's p50 end-to-end latency on our traffic is 4.4 seconds, which includes thinking.

• Context — both 1M tokens. Output ceiling differs: 131,072 for Mercury 3, 128,000 for Opus 5.5.

• Inputs — Opus 5.5 takes text, images and files; Mercury 3 takes text. This is not a minor line item. Any document, screenshot, diagram or PDF in your pipeline silently disqualifies Mercury 3.

• Reasoning control — Mercury 3 exposes instant, low, medium and high effort and defaults to low when the field is omitted. Opus 5.5 uses adaptive thinking with an effort dial of its own. Both default in a way that will surprise you if you do not set the field explicitly: Mercury 3 at low, Opus at whatever Anthropic's adaptive policy chooses for the request.

• Evidence — Opus 5.5 has an independent index score and a vendor benchmark table; Mercury 3 has neither a third-party score nor a numeric vendor score.

The one agent result Inception did publish

Buried under the throughput claims is the only Mercury 3 quality measurement with a defined task: a web-agent harness run on 50 live Mind2Web navigation tasks, in which Mercury 3 completed 50% and, Inception says, "significantly more than GPT-6.1-Sol", using the same token volume and number of agent steps.

Read that carefully. The comparison model is GPT-6.1 Sol, not Opus 5.5. The harness is Inception's own production one. The sample is 50 tasks. And "same token volume and same number of agent steps" is the important part for cost, because it means the per-task saving really is the per-token ratio rather than an artefact of the model taking fewer steps to get there. A 50% completion rate on live site navigation is a credible result for a fast model; it is not a claim that Mercury 3 would finish an Opus-grade coding task.

For Opus 5.5 the equivalent evidence is Terminal-Bench 4.0 at 59.6% on Artificial Analysis's harness — a different benchmark, a different task family, and not directly comparable. The honest statement is that nobody has run both models on the same agent benchmark and published the result. Until someone does, every claim about which is better at agent work is inference from positioning.

What a week of traffic costs on each

Take a workload that is plausible for both: 2 million input tokens and 500,000 output tokens per day, roughly a mid-sized support or research assistant with a few hundred thousand calls. On Mercury 3 at the launch rate that is $0.20 of input and $0.175 of output — under forty cents a day. At list price it is $0.40 and $0.35, or 75 cents. On Opus 5.5 the same traffic is $8.00 of input and $10.00 of output, or $18.00 a day.

Now add the part the token math hides. If Mercury 3 needs a second attempt on one call in five, and each retry costs the same as the original, the effective price rises by 20% — to about 45 cents a day. Even in the absurd case where every single call had to be repeated once, the bill would be 75 cents, still 24 times below Opus 5.5. The price argument is not close, and that is precisely why the decision has to be made on quality and capability instead.

Turn it around and the same arithmetic explains why people still pay for Opus. At $18 a day, one avoided escalation per day — a support turn that does not reach a human, a code change that does not need a second review — pays for the entire Opus bill several times over. The model is expensive per token because the work it replaces is more expensive still.

The capability lines that do not show up in a price table

Opus 5.5 accepts images and files, which means it can read the screenshot a user attaches, the PDF a contract lives in, or the diagram in a design doc. Mercury 3 is text in, text out. If your workload touches any of those, the comparison ends there and the price gap is irrelevant.

Opus 5.5 also carries Anthropic's tooling surface: structured outputs, tool use, a batch API at half price for asynchronous work, and prompt caching with explicit write and read rates. Mercury 3 has tool calling and structured outputs on an OpenAI-compatible endpoint, which is genuinely convenient — it drops into existing code without a rewrite — but its cache story is a discount on input rather than a documented cache-write lifecycle.

The inverse capability line is speed. Opus 5.5 at 75.6 tokens per second is fine for a coding agent working through a long task and painful for anything a person is waiting on. Mercury 3's pitch is that a person waiting on it does not notice the wait. If your product is a voice turn, an autocomplete, a search answer or a tutor's hint, that difference is the product, not a nice-to-have.

How to run both without betting on either

Claude Opus 5.5 is on our catalogue as anthropic/claude-opus-5.5 at Anthropic's list rates — $4 input, $20 output, $0.20 cache reads — with 0% markup, so a change to Anthropic's rate card shows up on our side the same day rather than at the next billing cycle. Mercury 3 is not on our catalogue: our model API returns no Inception models at all, and the honest instruction is to reach it through Inception's own platform. We are not going to pretend otherwise to make a routing pitch land.

What the split is good for is exactly this decision. Put Mercury 3 behind Inception's key on the latency-critical tier, keep Opus 5.5 as the escalation path behind one OpenAI-compatible endpoint alongside more than 200 other models, and let automatic failover catch the case where the new provider wobbles in its first month. The routing DSL can express the handoff — fast model first, escalate on a low-confidence signal or a tool-call failure — without a second integration for the fallback. You end up measuring Mercury 3 against your own traffic instead of against a launch post, and you keep the model that can read a screenshot for the requests that need it.

Which one to pick

If your workload is text-only, latency-bound, and tolerant of an occasional wrong answer that a human or a retry can catch, Mercury 3 at launch pricing is one of the cheapest credible options available this week, and the 100 million free tokens on Inception's platform make the trial close to free. Try it on your own traffic, pin the effort setting, and measure the retry rate rather than the throughput.

If your workload involves documents, images, long-horizon code changes, or answers that nobody checks before they are acted on, Opus 5.5 remains the defensible choice and the price is not the interesting variable. Its independent index score, its multimodal input and its published benchmark table are the things you are buying, and none of them has a Mercury 3 equivalent yet.

The mistake to avoid is treating the 57x as a migration mandate. It is a discount on tokens, and tokens are the cheapest part of most AI systems. The expensive parts — a human reviewing output, a wrong answer reaching a customer, a pipeline rebuilt around a model that cannot see — are exactly the parts the price table does not show.