Generated title card for Claude Opus 5.5 vs GLM-5.4, with three chips reading 'GLM-5.4: no such model', 'GLM-5.3 open weights, 18 Aug 2026' and 'Opus 5.5 GA 22 Sep 2026', and a footer noting Z.ai documentation and Hugging Face were checked 19 November 2026. OrcaRouter logo bottom-right.
Guides & Insights

Claude Opus 5.5 vs GLM-5.4: The Cheap Token Is Not the Cheap Task

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

The most expensive mistake in model selection is comparing price lists instead of finished tasks. Claude Opus 5.5 costs $4 per million input tokens and $20 per million output. GLM-5.3 — the model Z.ai actually ships, and the one anyone typing "GLM-5.4" is looking for — costs $1.40 and $4.40. On the sticker, the Chinese open-weights model is roughly a fifth of the price. On Artificial Analysis's own measurement of what a task costs to complete, the gap collapses to 2×. That collapse, not the price list, is the story of this comparison.

First, the name. GLM-5.4 does not exist. Z.ai's documentation index lists GLM-5.3 and GLM-5.3-Flash as the current text models, the lab's Hugging Face organisation has no repository with "5.4" in it, and a full-text search of the Hub on 19 November returns nothing. The number is a search term, not a release. GLM-5.3 shipped in August 2026 — 18 August by Artificial Analysis's release date, with the weights on the Hub from 25 August — and it is the model every claim below is about.

What GLM-5.3 is, and what it costs

GLM-5.3 is Z.ai's open-weights text flagship. It uses the same base model as GLM-5.2, with every gain coming from post-training, which is why the licence and the architecture are unchanged and why the improvements concentrate in exactly two places: complex coding and long-horizon agentic work. It has a 1M-token context, 128K max output, text-only input, and reasoning that cannot be switched off — three effort levels, low, high and max, defaulting to max.

Headless capture of Z.ai's developer documentation page for GLM-5.3, showing the model overview, the note that it uses the same base model as GLM-5.2 with all gains from post-training, and the feature section stating a 1M-token context window and 128K maximum output.

• Price — Claude Opus 5.5 $4.00 input / $20.00 output / $0.20 cache read vs GLM-5.3 $1.40 / $4.40.

• Weights — closed API only vs Apache-style open release on Hugging Face, downloadable and self-hostable.

• Context — 1M tokens both, 128K max output both.

• Effort control — Opus 5.5 runs adaptive thinking that cannot be disabled, default medium vs GLM-5.3 reasoning always on, default max.

• Intelligence Index — 59.1 vs 31.7 on the v5.0.1 revision, a 27.4-point gap.

• Cost per Index task — $10.04 vs $5.01, a 2× gap, not the 4.5× the sticker implies.

• Modality — text and image input vs text only.

Why a 4.5× price advantage becomes a 2× cost advantage

Artificial Analysis ran both models through the same Index suite and counted the tokens each one spent. Claude Opus 5.5 generated 218M output tokens across the suite. GLM-5.3 generated 135M. That is not the direction most people expect — the expensive model is the more verbose one here — but the ratio that matters is against the price, and GLM-5.3's 1.6× token advantage does not come close to covering its 4.5× per-token advantage in the other direction.

Generated two-column scoreboard titled 'Claude Opus 5.5 vs GLM-5.3 - the scoreboard'. Left column Claude Opus 5.5 rows: Intelligence Index 59.1, cost per task $10.04, Terminal-Bench 4.0 59.6%, FrontierCode 2.0 46.4%, weights closed, price $4 / $20. Right column GLM-5.3 rows: Intelligence Index 31.7, cost per task $5.01, Terminal-Bench 4.0 41.9%, FrontierCode 2.0 21.6%, weights Apache open, price $1.40 / $4.40. Footer: both per Artificial Analysis index v5.0.1, GLM-5.4 does not exist. OrcaRouter logo bottom-right.

Run the arithmetic the way a budget does. On a task where both models spend the same number of tokens, GLM-5.3 is 4.5× cheaper. In practice GLM-5.3 spends about 0.62 tokens for every one Opus 5.5 spends, so the effective ratio lands near 2.8× before input tokens, and the measured per-task figure of $5.01 against $10.04 puts it at 2×. Still a real saving — roughly half — but half is a different decision from a fifth. A team that budgeted on the sticker and got a 2× bill will have built the wrong business case.

There is a second-order effect worth flagging because it caught out more than one team: GLM-5.3's reasoning defaults to max. A migration that only swaps the model ID and leaves the reasoning default untouched gets the most expensive and slowest configuration of the model, which is precisely where the verbosity comes from. Z.ai's own migration note is explicit that disabling reasoning is no longer supported and that low must be set deliberately.

The capability gap is real, and it is not uniform

A 27-point index gap sounds disqualifying, and for general knowledge work it nearly is. But the per-benchmark numbers separate the two models along a line that matters more than the composite. All figures below are Artificial Analysis's, same harness, both at their highest effort setting.

• Terminal-Bench 4.0 — 59.6% vs 41.9%. An 18-point gap on agentic terminal work.

• FrontierCode-AA 2.0 — 46.4% vs 21.6%. Opus 5.5 more than doubles GLM-5.3.

• Terminal-Bench-Science 0.1 — 59.0% vs 9.5%. The largest gap on the board, and a rout.

• HLE-Diamond — 67.9% vs 17.9%.

• AA-Omniscience — 48.1 at 65.5% accuracy vs 6.0 at 32.3% accuracy.

• GDPval-AA v2.1 — 1866 Elo vs 1655, the narrowest gap of the set at 211 points.

• ITBench-SRE — 38.2% vs 46.1%, one of the few places GLM-5.3 wins.

• LCR (long-context reasoning) — 84.7% vs 79.7%.

The pattern is that GLM-5.3 holds up on long-context reasoning and on site-reliability tasks, and falls apart on science-heavy terminal work and knowledge recall. If your workload is document analysis over a long context, the 27-point composite overstates the difference. If it is agentic coding on unfamiliar repositories — which is what both vendors market — the composite understates it.

The vendor table Z.ai publishes compares Opus 4.8, not Opus 5.5

Worth knowing before you trust any GLM-versus-Claude number you find in a chart. Z.ai's GLM-5.3 model card carries a benchmark table with an "Opus 4.8" column and a "Fable 5 (w/ fallback)" column. Claude Opus 5.5, released five weeks after GLM-5.3, does not appear in it. That is not dishonesty — the table was built when 5.3 shipped and Opus 5.5 did not exist yet — but it does mean the comparison most people are searching for has never been published by either vendor, and the only same-harness numbers available are the independent ones above.

Z.ai's own numbers do carry one claim worth taking seriously, because it points at the open-weights model's genuine strength: on CyberGym, its vulnerability-discovery benchmark, GLM-5.3 scores 84.5 against Opus 4.8's 78.1. The card describes the gain as "emergent" — cyber capability scaled faster than the team expected as post-training scaled — and the exploitation benchmarks more than double from GLM-5.2. Anthrop​ic published an analysis of GLM-5.3's cyber capabilities on 29 September, which is not a thing vendors usually do about a competitor's open-weights release. If your work is defensive security tooling, that is the one axis where the open model is genuinely ahead, and it is also the axis where the open weights matter most: you can run the classifier on your own infrastructure rather than sending exploit-adjacent prompts to someone's API.

Where the open weights actually change the decision

Everything above is a capability and cost comparison, and on both counts Claude Opus 5.5 wins for most work. The reason GLM-5.3 still gets deployed is neither: it is that a 5.3-family model downloads, runs behind your own gateway, and never leaves your network.

Three concrete situations where that outweighs 27 index points. Regulated data that cannot cross a jurisdiction boundary, where the choice is GLM-5.3 or nothing. High-volume classification where a 300M-parameter fine-tune of the base beats the general model on your labels at a fraction of the per-call cost. And cyber-adjacent tooling, where the benchmark advantage and the self-hosting requirement point the same way.

For everything else the honest recommendation is to keep the routing decision cheap. Both models are on one OrcaRouter key — anthropic/claude-opus-5.5 and z-ai/glm-5.3 — at 0% markup on provider list price, which matters more for the Z.ai side than the Anthrop​ic side because Z.ai repriced this line twice in 2026. Route the high-volume, low-stakes passes to GLM-5.3 and the planning loop to Opus 5.5, and let failover cover the case where the cheap tier's provider has a bad afternoon. The routing DSL makes that a config change rather than two integrations, and the blended per-task figure is the one to put in the budget rather than either sticker.

Headless capture of OrcaRouter's model page for z-ai/glm-5.3, showing the model name and its sections for code samples, pricing, performance and public benchmarks.

The answer, with the name settled

Claude Opus 5.5 is the better model for agentic coding, science-heavy terminal work and anything where recall matters. It costs roughly twice as much per finished task as GLM-5.3, not four and a half times, and it is worth the difference for the workloads both vendors advertise.

GLM-5.3 is the right choice when the weights have to be yours, when the workload is long-context reasoning or on-call triage, or when the job is cyber-defence. It is a genuine frontier-adjacent open model at a genuinely low price, and its biggest weakness — science-terminal work at 9.5% — is narrow enough that a workload-specific evaluation will tell you in an afternoon whether it matters to you.

GLM-5.4, the model in the title, is not part of the decision. It is a version number with no release behind it, and if Z.ai ships it the arithmetic above is what it will have to beat.

Compared in this article3

Detected from this article · Benchmarks: Artificial Analysis · updated daily