Claude Mythos 5 vs GLM 5.5: One You Cannot Buy and One That Does Not Exist

This is the only matchup in the current batch where both sides are unobtainable, for two entirely different reasons. C​laude M​ythos 5 is real, dated, benchmarked and gated — A​nthropic released it on June 9, 2026 as the restricted sibling of Claude Fable 5, refreshed it as C​laude M​ythos 5.1 on September 1, and still hands it out through a verification program rather than a price list. G​LM 5.5 is a name with nothing behind it: as of today there is no model card, no API identifier, no pricing row, no Hugging Face checkpoint and no release note at Z.ai or anywhere else that uses it. So the honest version of this comparison is a substitution — Claude Fable 5, the ungated twin of C​laude M​ythos 5, against GLM-5.3, the flagship Z.ai actually shipped in August — plus a precise account of why one of these names is a gate and the other is a rumour.

Where "G​LM 5.5" comes from, and the five checks that settle it

Z.ai open-weights its flagships and publishes them on a predictable cadence, so a released model leaves the same five fingerprints every time. All five were re-checked today, and all five are negative.

• The Hugging Face organisation. The four newest repositories in zai-org are the GLM-5.3 series, all created 2026-08-25: GLM-5.3, GLM-5.3-BF16, GLM-5.3-Flash and GLM-5.3-Flash-BF16. The newest entry before those is GLM-5.2 from 2026-06-16. A repository search for "G​LM-5.5" returns two false positives — EXL2 quantisation filenames using 5.5 as a bits-per-weight value — and nothing else. A search for G​LM-5.4 returns nothing at all.

• The vendor's release notes. The newest entry on Z.ai's own release log is GLM-5.3-Flash, followed by GLM-5.3, GLM-5.2, G​LM-5.1 and G​LM-5. There is no 5.4 line and no 5.5 line.

• The vendor's documentation index. The model enumeration Z.ai publishes for tooling lists GLM-5.3-Flash, GLM-5.3-FlashX, GLM-5.3, GLM-5.2 and earlier. The 5.5 entry does not resolve.

• Our own catalogue. z-ai/glm-5.3 returns a live card with a full rate card and specifications. z-ai/glm-5.5 returns "model not found", and so does z-ai/glm-5.4.

• The independent board. Artificial Analysis carries a full GLM-5.3 entry — an Intelligence Index of 45 at its maximum setting, a measured cost of $2.01 per index task, throughput of 64.8 output tokens per second with a 3.3-second first token. There is no 5.5 entry to compare against.

Five surfaces, five negatives, one day. The name entered circulation attached to a release window rather than to a product, and the window closed without it. Analysts reading Z.ai's roughly 54-to-70-day flagship cadence now project the next release into a window running from early October to early November, and most of them expect it to be numbered 5.4. The safe assumptions if it does arrive are a 1M-token context, since GLM-5.2 and GLM-5.3 both carry one, and open weights, since that is Z.ai's track record — a pattern, not a promise.

The gate on the A​nthropic side is a different kind of nothing

C​laude M​ythos 5's problem is not existence. It is that A​nthropic sells every frontier generation twice from the same underlying weights — a M​ythos SKU with cyber and biology safeguards tuned down for vetted defenders, and a Fable SKU with the full safeguard stack left on — and only one of them is purchasable. C​laude M​ythos 5 was the restricted half of the June 9 generation; Claude Fable 5 was its public twin. On September 1 the pair repeated: C​laude M​ythos 5.1 restricted, Claude Fable 5.1 generally available.

There is a price on the restricted SKU — $10 per million input tokens and $50 per million output — and eligibility is a verification bar rather than a purchase, so for almost everyone the price is decorative. On October 6 A​nthropic rebuilt the Cyber Verification Program into three application-based tiers and folded its earlier partner program into it, which turned C​laude M​ythos 5.1 from an allocation into something a qualifying security organisation can apply for. That is real movement toward openness, and it is still an application review aimed at security organisations rather than a signup page.

So the shape of this matchup is: one side is a product you cannot buy, and the other is a product that does not exist. Neither is a model you can put in a config file today, and a comparison page that pretends otherwise is selling you a spec sheet it invented.

The comparison you can actually run

Swap in the model that exists and the matchup becomes real, close in one place and lopsided in another.

• Input price — Claude Fable 5 at $10.00 per million tokens vs GLM-5.3 at $1.26, per our live catalogue; GLM-5.3 is 87% cheaper.

• Output price — $50.00 per million vs $3.96; GLM-5.3 is 92% cheaper.

• Cache reads — $0.25 per million on Claude Fable 5.1, the newer SKU, vs $0.234 on GLM-5.3 — the one line where the A​nthropic model is not the expensive one.

• Context and output ceiling — 1M tokens and 128K maximum output on both.

• Modality — text, image and file input on the A​nthropic side; GLM-5.3 is text-only.

• Independent quality — GLM-5.3 at 45 on Artificial Analysis's Intelligence Index against Claude Fable 5.1 at 53 at its maximum setting, on the same board.

• Measured cost per index task — GLM-5.3 at $2.01 against Claude Fable 5.1 at $7.63. The list-price gap is roughly 12× on output; the realised gap is under 4×.

That last pair is the useful one, and it is the same pattern that shows up across every open-weights-versus-frontier comparison: the list price overstates the saving by a large factor, because the cheaper model burns more tokens to finish the same task. A budget built on the rate card alone will be wrong in the optimistic direction.

The other interesting row is the one Z.ai wrote about its own model. The catalogue card for GLM-5.3 describes it as delivering roughly a 50% improvement in coding experience over GLM-5.2 and — the sentence that belongs in this article — as matching M​ythos 5 on selected cybersecurity capabilities. That is the vendor whose 5.5 does not exist making a parity claim against the model this comparison is nominally about, using its shipped 5.3. It is a vendor claim, unreproduced, and it points at something true about the market: the restricted-tier capability that justified a $10 / $50 rate card two generations ago is now something an open-weights lab claims at a twentieth of the output price.

What open weights would change, if 5.5 ever arrives

If G​LM 5.5 ships as a normal Z.ai flagship it will almost certainly be an open-weights model, and that is the dimension on which this pairing would actually turn — more than any benchmark row. The price argument is already settled and it favours Z.ai heavily. The capability argument favours A​nthropic on the long-horizon agentic and multimodal work that enterprise buyers mostly run.

What open weights change is everything that is neither. Whether you can run it inside your own boundary. Whether you can fine-tune it on data you cannot send anywhere. Whether you can audit the thing you are depending on. Whether a vendor can deprecate it out from under you — and that last point is not hypothetical in this pairing. C​laude M​ythos 5's generation already had a two-week worldwide unavailability in June after export controls landed, restored to some US organisations on July 1. A model behind a verification program is a model whose access policy is somebody else's decision, and a downloaded checkpoint has no access policy at all.

The counterweight is that self-hosting a frontier-class open model is not free in the way a download button implies. The weights are large, the serving hardware is not commodity and the operations burden lands on you. Most buyers do not self-host; they rent the open model from a host at close to the vendor's own rate. The practical benefit of open weights for most teams is optionality, and optionality is worth paying for only if you would exercise it.

Getting either capability behind one key today

Neither C​laude M​ythos 5 nor G​LM 5.5 is something a platform can hand you: one is gated by its vendor and the other has no artifact to serve. What is routable is the reachable member of each family, and that is where the practical decision lives. OrcaRouter carries anthropic/claude-fable-5.1 at A​nthropic's list price of $10.00 per million input and $50.00 per million output with the reduced $0.25 cache read, and z-ai/glm-5.3 at $1.26 and $3.96 — both behind the same O​penAI-compatible endpoint, provider list price passed through at 0% markup, with automatic failover across providers when one route degrades.

Two reasons that matters more in this pairing than in most. The first is that the Z.ai side of the comparison has already moved its prices twice this year and its model line three times, so a pass-through route that mirrors the vendor's card beats a number frozen into a contract. The second is that the honest recommendation here is not to choose: the cheap model takes the high-volume text work and the expensive one takes anything with an image in it or a multi-step plan behind it, and that split is a routing configuration rather than two procurement conversations. The routing DSL expresses it in one call — cheap path by default, escalate on a threshold — and model fusion is available for the narrow case where a second opinion is worth more than the tokens it costs.

If you are waiting for G​LM 5.5

Five artifacts, in this order, and none of them is a rumour: a model card with a stated parameter count; an API identifier in Z.ai's documentation; a pricing row on the vendor's own rate card; a checkpoint in the zai-org Hugging Face organisation with a licence file attached; and a release note on Z.ai's own site. Watching for those is cheaper than re-reading analyst forecasts, because the forecasts have already been wrong once about this exact name.

In the meantime the discipline is to benchmark against something with weights. GLM-5.3 is the model to measure, and Claude Fable 5.1 is the model to measure it against if the comparison you care about is open-weights efficiency versus frontier capability. Comparing against a name that does not resolve is not a head start on the evaluation; it is an evaluation you will have to run again.

None of that is hypothetical: 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.