Claude Fable 5 takes the quality crown at $20/M, GLM 5.2 narrows the open-source gap
Anthropic's Fable 5 hits 59.9 quality at $20/M. GLM 5.2 reaches 51.1 open-source at $1.46/M. Opus 4.8 replaces 4.7 at same price.
Claude Fable 5 is the new quality leader — but at what cost?
Claude Fable 5 (Anthropic) lands at 59.9 on the quality index, the highest score in the market right now. It costs $20/M input tokens — nearly double Claude Opus 4.8 at $10/M and almost 1.8x GPT-5.5 at $11.25/M. At 77 tokens per second, it's faster than both Opus variants but trails GPT-5.5's 83 tok/s.
The operational question: does 4.2 points of quality over Opus 4.8 justify 2x the price? For pipelines where output quality gates human review cycles, yes — fewer review passes can offset the token premium. For batch workloads or high-volume inference, the math breaks fast.
Opus 4.8 quietly replaces 4.7 at the same price
Claude Opus 4.8 scores 55.7 at $10/M, up 2.2 points from Opus 4.7 at 53.5. Same price, same tier, modest speed bump (58 tok/s vs 52 tok/s). If you were running Opus 4.7 in production, switching is a free upgrade — no cost delta, no latency regression.
Against GPT-5.5 (54.8 at $11.25/M), Opus 4.8 now wins on both quality and price. The gap is small (0.9 points), but it inverts the value comparison from last week.
| Model | Quality | Price/1M | Speed |
|---|---|---|---|
| Claude Fable 5 | 59.9 | $20.00 | 77 tok/s |
| Claude Opus 4.8 | 55.7 | $10.00 | 58 tok/s |
| GPT-5.5 | 54.8 | $11.25 | 83 tok/s |
| GLM 5.2 | 51.1 | $1.46 | 139 tok/s |
GLM 5.2 is the open-source story this week
GLM 5.2 (Z AI) hits 51.1 quality at $1.46/M — open weights, and it outperforms GPT-5.4 (51.4) which costs 3.9x more. At 139 tok/s, it's also faster than every closed model above it except Gemini 3.5 Flash. For teams that need self-hosting or want cost predictability without API lock-in, this is the strongest open-source value proposition in the table.
DeepSeek V4 Pro remains the budget open-source pick at 44.3 quality and $0.54/M. But GLM 5.2 closes the quality gap to 6.8 points for $0.92/M more — a reasonable trade for production workloads where parser failures and retry costs dominate.
What to watch
- Whether Fable 5's $20/M holds or Anthropic tiers it down as Opus 4.8 cannibalizes the mid-premium segment.
- GLM 5.2 self-hosted benchmarks vs API numbers — open-source claims need independent confirmation on real workloads.
- GPT-5.5's response: Opus 4.8 now beats it on both price and quality, which hasn't happened before in this tier.
For a full comparison across all tiers, see LLM Selector or Explore the leaderboard.
Stay in the loop
Weekly LLM analysis delivered to your inbox. No spam.