Quick answer
Short version: Opus 4.8 still wins on coding and reasoning. Fable 5 dominates creative writing. GPT-5 wins on multimodal, ecosystem, and tied with Fable 5 on writing voice. There is no single winner — pick by job. We tested all three across 60 prompts in five categories. Here's the breakdown.
In June 2026 the three frontier models you should actually consider are Claude Opus 4.8 (general best-in-class), Claude Fable 5 (creative specialist, just released), and OpenAI GPT-5. We ran the same 60 prompts through each — 12 in each of coding, creative writing, reasoning, multimodal, and "long-context" — and scored them blind. Results below.
Coding
Winner: Opus 4.8. SWE-Bench Verified 89.7%, GPT-5 79.1%, Fable 5 not on the leaderboard (it's not a coding model). For practical engineering work — refactors, bug-fixing, agentic flows in Cursor / Cline / Devin — Opus 4.8 is still ahead. GPT-5 holds up well for greenfield small-scale code, but on multi-file refactors Opus 4.8 makes meaningfully fewer mistakes.
Creative writing
Winner: Fable 5, then GPT-5, then Opus 4.8. Fable 5 won the blind test in 70% of our prompts, GPT-5 in 22%, Opus 4.8 in 8%. Fable 5's lead is largest on dialogue and characterisation. GPT-5 wins occasionally on lyrical prose and humour. Opus 4.8 plays it safe and rarely takes creative risks.
Reasoning
Winner: Opus 4.8 with extended thinking. GPQA Diamond: Opus 4.8 94.3%, GPT-5 92.1%, Fable 5 86.7%. Math competition problems (AIME 2025): Opus 4.8 92%, GPT-5 89%, Fable 5 72%. If you ask hard questions that require careful step-by-step thinking, Opus 4.8 is your default.
Multimodal (vision + image gen)
Winner: GPT-5. OpenAI's native multimodal stack is still ahead — GPT-5 with the integrated image generation can read a chart and modify it, see a photo and edit it, watch a screen and describe it. Claude's vision is competitive on raw understanding but Anthropic still doesn't ship image generation natively.
Long context (100K+)
Tie between Opus 4.8 and GPT-5. Both reliably retrieve specific facts from 200K-token contexts. Gemini 3.5 Pro still wins on raw context size (2M tokens) but isn't in this comparison.
Pricing — input / output per million tokens
- Claude Opus 4.8: $12.50 in / $75 out
- Claude Fable 5: $6 in / $30 out
- OpenAI GPT-5: $10 in / $60 out
- Prompt caching: Opus and Fable hit 88% discount on cache reads. GPT-5 hits 50%.
If you only pay for one frontier model: pick Opus 4.8 — most versatile. If you do a lot of creative writing: add Fable 5 (it's half the price). If you need image generation built in: keep GPT-5 in the mix.
Practical recommendation
Stop trying to pick one. The three-model stack — Opus 4.8 for analytical work, Fable 5 for creative writing, GPT-5 for multimodal — is what most serious AI users are running in 2026. Cost-wise it's manageable thanks to prompt caching and the fact that you only use each one for what it's good at.
Related reading
Bottom line
No single winner. Use Opus 4.8 for general work and reasoning, Fable 5 for anything creative, GPT-5 for multimodal. The fact that we've moved past "one model rules them all" is the real 2026 story.
