Mercury 3 vs GLM-5.5: The Comparison the Search Query Asks For, Run Against the Model That Exists

There is no GLM-5.5. Z.ai's release log stops at GLM-5.3-Flash on August 26, 2026, its documentation index enumerates GLM-5.3, GLM-5.3-Flash and GLM-5.2, the Hugging Face zai-org organisation's newest four repositories are the GLM-5.3 series created August 25, Artificial Analysis has no page under that name, and our own catalogue returns model-not-found for z-ai/glm-5.5. Five surfaces, checked again today, all negative. So a page titled "Mercury 3 vs GLM-5.5" can do one of two things: invent a spec sheet for a model that has never existed, or tell you what the name probably refers to and run the comparison that can actually be executed. This page does the second. 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.

The model that exists is GLM-5.3, which Z.ai released on August 18, 2026 with open weights following on August 25. Beside it is Mercury 3, which Inception Labs released on October 19, 2026 as a diffusion language model with a 1M-token context and a 1,068 token-per-second throughput claim. That is a real matchup with a real question in it: a proven open-weights coding model at $1.26/$3.96 against an unmeasured speed model at $0.35 output during its launch window. The version number in the title is wrong. The comparison underneath it is not.

Why the name exists at all

GLM-5.5 is a projection, not an announcement. It traces back to an analyst forecast circulated in June 2026 that put Z.ai's next flagship in an August window; August closed with GLM-5.3 instead. Analysts reading Z.ai's cadence now expect the next release numbered 5.4, inside a window running roughly October 8 to November 9. Today is October 20 — inside that window, with nothing shipped.

This blog has covered the name a dozen times, always with the same finding, and the temptation with a new Mercury launch is to treat the missing opponent as permission to skip the article. It is not: the query is real, and the useful answer is a comparison against the model Z.ai actually sells. What follows lists what would have to appear before the title becomes accurate, then runs the numbers that exist.

Five things that would make GLM-5.5 a model

• An API identifier — z-ai/glm-5.5 resolving on any public catalogue. Ours returns model-not-found, as does glm-5.4.

• A rate-card row — a price on z.ai's model API page. The page currently lists GLM-5.3 ($1.40/$4.40), GLM-5.3-Flash ($0.15/$0.50), GLM-5.3-FlashX ($0.37/$1.25) and earlier models, and nothing above 5.3.

• A release-note entry — docs.z.ai's New Released page is dated by entry and tops out at 2026-08-26.

• A weights repository or an explicit no-weights statement — the zai-org feed's newest text repos are all dated 2026-08-25.

• A dated independent card — artificialanalysis.ai/models/glm-5-5 returns 404 while glm-5-3 returns a full page.

Any one of those would be a strong signal. All five appearing together is what a release looks like; none appearing is what a forecast looks like. The check is re-runnable, and it should be re-run rather than trusted, because a negative finding decays the moment the vendor ships.

The actual spec contrast

• Price — Mercury 3 $0.10/$0.35 per million at launch, $0.20/$0.70 at list. GLM-5.3 $1.26/$3.96 on our catalogue, which is 10% below Z.ai's own $1.40/$4.40 list.

• Context — both 1M tokens. Mercury 3 carries a 131,072-token output ceiling; GLM-5.3 caps output at 128,000.

• Architecture — Mercury 3 is a diffusion model, generating tokens in parallel; Inception says it is more than 40x larger than Mercury 1 without giving an absolute parameter count. GLM-5.3 is a mixture-of-experts transformer whose published detail Z.ai keeps at the family level.

• Speed — Mercury 3 at 1,068 tok/s, Inception's number, unreproduced. GLM-5.3 at 88.4 tok/s output measured across our own playground traffic over seven days, with a median end-to-end latency of 3.3 seconds.

• Independent score — GLM-5.3 scores 44.8 on Artificial Analysis's Intelligence Index and 74.8 on its Coding index, ranking 14th on coding among the models it tracks. Mercury 3 has no Artificial Analysis page.

• Task scores — GLM-5.3 posts 83.9% on Terminal-Bench 2.1 and 41.9% on Terminal-Bench 4.0, 91.7% on GPQA Diamond and 50.3% on τ²-Banking, all per Artificial Analysis. Mercury 3 publishes no numeric task scores; its only structured result is 50% completion on 50 live Mind2Web navigation tasks.

• Weights — GLM-5.3 is open, with checkpoints on Hugging Face from August 25. Mercury 3's weights are not published and Inception has made no commitment to publish them.

• Modalities — text in, text out on both.

The one line where GLM-5.3 is unambiguously the stronger buy is the cheapest to verify: it has been measured. A 74.8 coding index and an 83.9 Terminal-Bench 2.1 score tell you what the model does on tasks you can name. Mercury 3 tells you it beats Claude Haiku 5.5 without saying by how much, on what, or at which of its four effort settings.

What 1,068 tokens per second is worth, and to whom

Mercury's speed case is not that it finishes a coding task faster in wall-clock terms — a model that takes more steps or needs a retry can lose that race at any throughput. The case is that individual turns come back while a person is still looking at the screen. Inception's own customer quotes make this explicit: tutoring hints in about a second, search generation at a median 0.9 seconds, support replies that stay conversational.

GLM-5.3 at 88.4 tokens per second is a normal autoregressive speed and would be unpleasant for any of those. It is entirely adequate for an agent that runs for six minutes and reports back, which is the shape of workload its benchmark profile describes — repo-scale coding, multi-step engineering, tool use over long horizons.

That is the real fault line between these two models, and it is not price. It is whether a human is waiting. If yes, Mercury 3's throughput is the product and its missing benchmarks are a risk you can bound with a retry budget. If no, GLM-5.3's measured competence is worth the 11-fold output-price premium, because a wrong answer from a background agent costs more than the tokens it saved.

Cost on a coding-agent workload

Coding agents are input-heavy: a long repository context, tool definitions, and conversation history on every step. Take 300 million input tokens and 12 million output tokens a day, which is a small team running agents continuously. On Mercury 3 at the launch rate that is $30 of input and $4.20 of output, or $34.20. At list price, $60 and $8.40, or $68.40.

On GLM-5.3 at our catalogue rate the same traffic is $378 of input and $47.52 of output, or $425.52 a day. Z.ai's list price would be $420 and $52.80, or $472.80. So the premium is roughly 6.2 times at Mercury's list price and 12.4 times during the launch window.

Now the honest counterweight: if Mercury 3's completion rate on a real coding task is materially below GLM-5.3's, that multiple collapses. A model that finishes two thirds of tasks needs 50% more attempts, and every attempt re-sends the repository context. At a 33% shortfall the effective premium falls to about 4 times — still large, still real, and no longer a decision you can make without measuring. The Mind2Web result says Mercury can navigate a live website; it says nothing about whether it can land a patch.

The open-weights question, which is not a pricing question

GLM-5.3's weights are on Hugging Face, which means the model can be self-hosted, fine-tuned on your own codebase, or pinned against deprecation. For a team whose competitive advantage lives in a fine-tune, that is the entire argument, and no hosted price competes with it. Mercury 3 offers no such path: it is an API product, and the announcement's enterprise tier is about dedicated capacity and data retention rather than about handing you the weights.

The counterargument is economic. Serving a model of GLM-5.3's class yourself costs more in GPUs and ops than $1.26 per million input tokens for most teams, and the open licence is insurance rather than a plan. It matters when the vendor changes terms, and it does not matter on a Tuesday.

Running the two side by side

GLM-5.3 is on our catalogue as z-ai/glm-5.3 at $1.26/$3.96 — the provider's list rate with 0% markup, so when Z.ai moves its price the change is live on our side the same day rather than at the next billing cycle. Mercury 3 is not on our catalogue; our model API returns model-not-found for every Inception model, so the route to it is Inception's own platform, which hands new accounts 100 million free tokens. We route GLM-5.3 and we do not route Mercury 3, and the difference in how you reach each one follows from that rather than from a preference.

The way to settle this matchup with your own traffic rather than with a launch post is a blind routing split: send half your agent's steps to Mercury 3 through Inception's key at a pinned effort level, half to z-ai/glm-5.3, and compare completion rate and total cost per finished task rather than per token. One OpenAI-compatible endpoint covers the GLM side along with more than 200 other models, with automatic failover, so if the new provider's first month is uneven your fallback is already wired. The routing DSL can hold the split as a rule so it is a configuration change rather than a code change when the answer comes back.

What to do while GLM-5.5 remains a name

If you arrived here looking for GLM-5.5, the truthful answer is that Z.ai has not shipped it and the number is a forecast. Watch the release-notes page and the Hugging Face feed; when a repository appears, the name becomes a model and this comparison becomes obsolete in the good way. Analysts expect the next flagship in the October 8 – November 9 window, most of them numbered 5.4, and today sits inside that window with nothing published.

If you need to decide today, the choice is between a measured open-weights coding model at roughly six to twelve times the price and a much faster text model with no published benchmark scores. Pick GLM-5.3 for work whose output you will not read before it ships. Pick Mercury 3 for work where a person is waiting and a retry is cheap. And budget Mercury 3 at its list price, because the discount that makes it look free is the one number Inception has not promised to keep.