← All posts

Since Z.ai released GLM-5.2 on 13 June, everyone has been amazed by the benchmarks and the price. Can you trust the numbers? Yes and no. The scores are real and you can use Z.ai's API directly today — but most enterprises will need to self-host for data-residency and compliance reasons, and priced that way the true numbers stop matching the headlines. In our own testing, GLM-5.2 also lands short of both Claude and OpenAI's Codex.

The short answer: Claude Fable 5 remains the quality ceiling for complex coding work — 90 percent on FrontierSWE against GLM-5.2's 74.4 — at $10/$50 per million tokens. GLM-5.2 is so cheap in comparison that it is a serious contender for high-volume workloads where you can justify the per-hour cost of renting GPUs, and with no managed GLM-5.2 on the hyperscalers at all (AWS Bedrock lists GLM 5, and Microsoft Foundry's Fireworks listing only reaches GLM 5.1), self-hosting is usually how you would run it. GPT-5.5 trails on these benchmarks but remains the most widely deployed of the three.

In practical terms you need to do three things: test on your own repositories before trusting any leaderboard, choose a deployment route deliberately, and price the models per completed task rather than per token. This piece works through each. The governance decision — API, self-hosted weights, or neither — is in our decision guide to open-weight frontier models; the UK compliance mechanics are in our data-residency brief.


Read the Benchmark Numbers Sceptically

Most published SWE-bench Pro scores are vendor self-reported figures from tuned agent scaffolds. Scale AI's standardised leaderboard, which runs every model on identical scaffolding, is humbler: per the SWE-bench Pro tracking, the vendor aggregate credits Opus 4.8 with 69.2 percent while Scale's best standardised Claude run scored 51.9 on the public set. A score 10 to 30 points above the standardised leaderboard is a scaffold number, not a model number. Fable 5's 80.3 on the same benchmark is likewise vendor-reported.

GLM-5.2's 62.1 has its own caveat: it was measured by third parties, as Z.ai published no SWE-bench figure at launch. The honest reading is narrower than the headlines: GLM-5.2 beats GPT-5.5 on like-for-like runs and sits within a point of Opus 4.8 on FrontierSWE, where the scores come from one analysis. The 69.2-versus-62.1 comparison crosses harnesses and should not be quoted as a seven-point gap.

FrontierSWE scores across four models Four horizontal bars on a 0 to 100 scale. Fable 5 leads at 90; Opus 4.8, GLM-5.2 and GPT-5.5 sit within 2.5 points of each other in the mid-70s. Fable 5 Opus 4.8 GLM-5.2 GPT-5.5 Claude Fable 5 — FrontierSWE: 90.0 Claude Opus 4.8 — FrontierSWE: 75.1 GLM-5.2 — FrontierSWE: 74.4 GPT-5.5 — FrontierSWE: 72.6 90.0 75.1 74.4 72.6 0 20 40 60 80 100
FrontierSWE, percent. GLM-5.2 sits level with Opus 4.8; Fable 5 is the outlier. Sources: BenchLM and VentureBeat, June 2026.

The practical rule: trust orderings within a single standardised run, treat cross-source comparisons as directional, and pilot on your own repositories before believing any of it.


What Each Model Offers an Engineering Organisation

GLM-5.2: frontier capability at commodity price

The economics are the headline: $1.40 per million input tokens and $4.40 output, with cached input at $0.26, against $5/$25 for Opus 4.8 (llm-stats comparison). For token-heavy agentic workloads — long refactors, whole-repository analysis using the 1M-token context — the gap compounds daily.

Switching costs are the under-reported fact. Z.ai exposes an Anthropic-compatible endpoint, so GLM-5.2 drops into Claude Code, Cline, and Cursor by changing a base URL and key — in Claude Code the 1M-context variant is glm-5.2[1m] (setup guide). Teams with agentic tooling built around Claude, as in our coding tool comparison, can trial it in the same workflow in an afternoon. The compatibility also means adopting GLM-5.2 creates no lock-in.

The weaknesses mirror the price. No vendor with a UK legal presence stands behind the output, the benchmark record is thinner, and the data-residency analysis must precede any pilot on real code. The managed-cloud route is not built yet: at the time of writing, AWS Bedrock's catalogue carries GLM 5 rather than 5.2, and Microsoft Foundry's Fireworks listing only reaches GLM 5.1 — so procurement cannot simply tick the usual hyperscaler box for 5.2. Early adopters also report migration friction: model-id conventions differ by tool, and routing through intermediaries has produced tool-call errors in Cursor.

Claude Fable 5: the quality ceiling, priced accordingly

Fable 5, the first of Anthropic's Claude 5 family, leads FrontierSWE at 90.0 percent — some 15 points clear of everything else in this comparison (BenchLM) — and is the strongest choice where correctness on complex, multi-file work dominates cost. The ecosystem compounds it: Claude Code holds a 46 percent "most loved" rating in The Pragmatic Engineer's survey of 906 developers, and tool and model coming from the same lab shows up in community reports on output quality.

The premium is steep: $10/$50 per million tokens, double Opus 4.8, which stays available at $5/$25 as Anthropic's own mid-tier (batch processing halves Fable to $5/$25). The economical pattern is a split estate: Fable for planning, architecture, and the hardest debugging; a cheaper model for volume. Until June the volume tier meant Haiku- or Sonnet-class models. GLM-5.2 arriving at Opus-class quality changes what the volume tier can be.

One caveat without precedent: on 12 June a US export-control directive forced Anthropic to suspend Fable 5 for all customers; access returned on 30 June when the controls were lifted. Two weeks of a frontier model switched off by government order is a concentration-risk data point for any single-vendor estate — whichever jurisdiction it sits in.

GPT-5.5: behind on these benchmarks, ahead on distribution

GPT-5.5 trails both here — 58.6 on SWE-bench Pro, 72.6 percent on FrontierSWE. Coding benchmarks are not the whole of engineering work, though: its strength is sustained multi-step execution, as we covered in our GPT-5.5 analysis, and it is the default model in more enterprise tooling than either rival. It is not the budget option: at $5/$30 per million tokens its output costs more than Opus 4.8's, so the case is the ecosystem, not the price. For organisations standardised on OpenAI through Microsoft, the question is not whether GPT-5.5 tops a leaderboard but whether the gap justifies leaving an integrated estate.


The Comparison at a Glance

GLM-5.2 Claude Fable 5 Claude Opus 4.8 GPT-5.5
FrontierSWE 74.4% 90.0% 75.1% 72.6%
SWE-bench Pro 62.1 (third-party run) 80.3 (vendor scaffold) 69.2 (vendor scaffold — not comparable) 58.6
Price per MTok (in/out) $1.40 / $4.40 ($0.26 cached) $10 / $50 (~$1 cached; batch $5/$25) $5 / $25 ($0.50 cached) $5 / $30 ($0.50 cached)
Context window 1M tokens 1M tokens 1M tokens 1M listed; 2x input price above 272K
Access API, coding plans, MIT weights; no first-party hyperscaler offering yet API and vendor tools only API and vendor tools only API and vendor tools only
Vendor accountability Chinese jurisdiction; DPA offered US/UK contractual presence US/UK contractual presence US/UK contractual presence

Cost per Token Is Not Cost per Task

The price gap is per token; engineering organisations buy completed work. If a cheaper model needs two attempts where a stronger one needs one, the gap halves before you count the engineer's time reviewing the failure. On FrontierSWE, GLM-5.2 sits 0.7 points behind Opus 4.8 and 15.6 behind Fable 5; on your codebase the gaps may differ, and no public retry-adjusted cost data exists yet. A two-week pilot measures it: a modest real-world quality discount still leaves GLM-5.2 far cheaper per completed task; a large one erases the saving.

Price per million tokens across four models Horizontal bars on a 0 to 50 dollar scale. Fable 5 output costs $50, GPT-5.5 $30, Opus 4.8 $25, GLM-5.2 $4.40. INPUT, $/MTOK OUTPUT, $/MTOK GLM-5.2 Opus 4.8 GPT-5.5 Fable 5 GLM-5.2 Opus 4.8 GPT-5.5 Fable 5 GLM-5.2 — input: $1.40 per million tokens Claude Opus 4.8 — input: $5.00 per million tokens GPT-5.5 — input: $5.00 per million tokens Claude Fable 5 — input: $10.00 per million tokens GLM-5.2 — output: $4.40 per million tokens Claude Opus 4.8 — output: $25.00 per million tokens GPT-5.5 — output: $30.00 per million tokens Claude Fable 5 — output: $50.00 per million tokens $1.40 $5.00 $5.00 $10.00 $4.40 $25.00 $30.00 $50.00 $0 $10 $20 $30 $40 $50
List prices per million tokens. Cached input: GLM-5.2 $0.26, Opus 4.8 $0.50, GPT-5.5 $0.50, Fable 5 about $1. Fable 5 batch runs $5/$25. Sources: llm-stats, OpenAI's published price table and Anthropic pricing docs, June 2026.

Caching shifts the numbers, not the ordering. Agentic coding re-reads the same files continuously, and all four discount cached input: Fable 5 at about $1 (90 percent off base), Opus 4.8 and GPT-5.5 at $0.50, GLM-5.2 at $0.26. Caching does not change the ordering, and GLM-5.2 keeps its edge cached or not. Real workloads run well below list price across the board, as we noted for Claude Code in our tool comparison.


Which Model for Which Team


New posts land here first. Sign up for email updates.

Frequently asked questions

Is GLM-5.2 better than Claude for coding?

No on published benchmarks. Claude Fable 5 leads FrontierSWE at 90.0 percent to GLM-5.2's 74.4, and GLM-5.2 also trails Claude Opus 4.8 by 0.7 points on the same tracking. What GLM-5.2 changes is the price of near-Opus quality: $1.40/$4.40 per million tokens against $10/$50 for Fable 5 and $5/$25 for Opus 4.8. For many teams the practical answer is a split estate — Claude for the hardest work, GLM-5.2 for volume workloads where the per-hour cost of GPUs or the per-token API price can be justified.

Can GLM-5.2 be used with Claude Code or Cursor?

Yes. Z.ai exposes an Anthropic-compatible API endpoint specifically for coding tools, so GLM-5.2 drops into Claude Code, Cline, Cursor, and similar agent harnesses by changing the base URL and API key rather than the tooling. In Claude Code the 1M-context variant uses the model id glm-5.2[1m]. This compatibility is a large part of the enterprise story: the switching cost between GLM-5.2 and Claude models is a configuration change, not a workflow migration — in both directions.

Why do GLM-5.2 benchmark scores differ between sources?

Because the scores are produced on different scaffolds. Most published SWE-bench Pro figures are vendor self-reported numbers using tuned agent harnesses; Scale AI's standardised leaderboard, which runs every model on identical scaffolding, produces substantially lower scores across the board. GLM-5.2's widely cited 62.1 was measured by third parties, since Z.ai published no SWE-bench number at launch, while Claude Opus 4.8's 69.2 and Claude Fable 5's 80.3 are vendor figures. Treat cross-model comparisons as directional unless the scores come from the same standardised run.

How much cheaper is GLM-5.2 than Claude?

On list prices, GLM-5.2 costs $1.40 per million input tokens and $4.40 per million output. Claude Fable 5 costs $10 and $50 — roughly 7x more on input and 11x on output — and Claude Opus 4.8 costs $5 and $25. Cached input widens the gap further: $0.26 for GLM-5.2 against about $1 for Fable 5 and $0.50 for Opus. Per-token price is not per-task cost, though: if a cheaper model needs more attempts on hard problems the gap narrows, and self-hosting GLM-5.2 replaces the per-token meter with a GPU-hour bill. Validate on your own workload rather than a pricing page.

← All posts