Prices verified 2026-08-04; GLM 5.2 figures as of August 2026
Fable 5 pricing is pulled from Anthropic's published rates (LAST_VERIFIED 2026-08-04). GLM 5.2 rates are Zhipu list prices as of August 2026 and may change. Benchmarks are vendor-reported unless noted.
See all Fable 5 alternativesFable 5 vs GLM 5.2
Claude Fable 5 vs GLM 5.2
An honest, dated comparison of Anthropic's frontier coding model and Zhipu's open-weights challenger — by specs, benchmarks, price, and when each one is the right call.
TL;DR
GLM 5.2 is the cheapest open-weights competitor with a real coding score and an MIT license you can self-host. Fable 5 wins on agentic coding (SWE-bench Pro 80.3% vs 62.1%), safety, and the Anthropic ecosystem; GLM 5.2 wins on price (~7x cheaper input), self-hosting and an open license. Pick by whether you need Anthropic guarantees or open-weights control.
Specs at a glance
Fable 5 prices are Anthropic list rates; GLM 5.2 rates are Zhipu figures as of August 2026.
| Spec | Claude Fable 5 | GLM 5.2 |
|---|---|---|
| Model ID | claude-fable-5 | GLM-5.2 |
| Input price (per M tokens) | $10.00 | $1.40 |
| Output price (per M tokens) | $50.00 | $4.40 |
| Context window | 1M | 1M |
| Max output | 128K | ~131K |
| Open-source license | Closed API (proprietary) | Open weights — MIT |
| Data retention | 30-day mandatory | None (self-hostable) |
Fable 5 carries a mandatory 30-day data retention requirement with no zero-data-retention option; GLM 5.2 is open-weights (MIT), so you control where and how it runs.
Benchmark comparison
Fable 5 leads on the benchmarks that matter for agentic coding; GLM 5.2's best score is vendor-reported.
| Spec | Claude Fable 5 | GLM 5.2 |
|---|---|---|
| SWE-bench Pro (agentic coding) | 80.3% | 62.1% |
| FrontierCode (Diamond) | 29.3% | Not published |
| GDPval (knowledge work) | 1932 | Not published |
| OSWorld-Verified (computer use) | 85.0% | Not published |
Sources: Fable 5 figures from Anthropic's June 9, 2026 announcement. GLM 5.2's SWE-bench Pro 62.1% is vendor-reported by Zhipu. GLM 5.2 has not published FrontierCode, GDPval or OSWorld scores — absence of evidence, not evidence of absence.
How to decide
Route by what you actually need, not by the headline.
You need to self-host
GLM 5.2's MIT open weights let you run it locally, air-gapped, or fine-tuned — Fable 5 is a closed API you can only call.
You're cost-sensitive at scale
At ~$1.40/$4.40 vs Fable 5's $10/$50, GLM 5.2 is roughly 7x cheaper on input. For high-volume production traffic that clears both quality bars, GLM 5.2 wins on cost.
You need frontier agentic coding
Fable 5's 80.3% SWE-bench Pro and 29.3% FrontierCode are the strongest published scores; GLM 5.2 trails at 62.1% (vendor-reported) and publishes no FrontierCode figure.
Retention or export control matters
Fable 5's mandatory 30-day retention and lack of zero-data-retention can be a compliance blocker; Chinese models like GLM 5.2 also face US export-control and availability considerations. Fable 5's Anthropic checks win if your organization needs those guarantees.
You want the closest thing to frontier
If the task is hard enough that a wrong answer is expensive, Fable 5 is the closer-to-frontier choice on the published evidence.
Gap-to-price math: input is ~7x cheaper on GLM 5.2, but on frontier-hard agentic tasks a cheaper model's failed attempts aren't free. Run your own workload in the cost calculator before deciding.
Estimate the real cost
Run both models through your own token mix — the 7x price gap narrows on cache-heavy agent workloads.