Suraj Khaitan's routing playbook for Claude's 2026 model fleet
After a month of running all of Anthropic's 2026 models inside Claude Code on real codebases, Suraj Khaitan argues that model routing — not prompting — is now the highest-leverage skill for developers.
Score breakdown
The post frames model routing — not prompting — as the primary productivity lever in Claude Code, and provides a concrete cost-and-capability breakdown of five distinct models that developers must now choose between for every task.
- 01Anthropic's 2026 model ladder runs: Haiku → Sonnet 4.6 → Opus 4.8 → Fable 5 → Mythos 5
- 02Sonnet 4.6 is priced at $3/$15 per million tokens with a 1M-token context window
- 03Opus 4.8 is priced at $5/$25 per million tokens ($10/$50 in fast mode) and supports dynamic workflows with hundreds of parallel subagents
Suraj Khaitan's post describes a month-long experiment running Anthropic's entire 2026 model lineup through Claude Code on real engineering work — bug fixes, migrations, greenfield features, and test suites. His central thesis is that the highest-leverage skill in 2026 is no longer prompting but routing: assigning the right model to the right task. The lineup he describes forms a graded ladder — Haiku for high-volume, latency-sensitive glue work; Sonnet 4.6 as the everyday workhorse at $3/$15 per million tokens with a 1M-token context window; Opus 4.8 as the "heavy lifter" for architecture and judgment-heavy refactors at $5/$25 (or $10/$50 in fast mode); and Fable 5 as the frontier Mythos-class model for long-horizon coding, vision, and research at $10/$50. Mythos 5 shares the same underlying model as Fable 5 but has its safety classifiers removed and is restricted to vetted cyber-defense and biology partners at the same price point.
He also highlights Opus 4.8's standout improvement as honesty rather than raw capability — describing it as roughly four times less likely than its predecessor to let flaws in its own code pass unremarked.
A key distinction Khaitan draws is between Fable and Mythos: they are the same base model, with Fable shipping classifiers that hand off sensitive cyber/bio/chemistry queries to Opus 4.8, while Mythos has those guardrails lifted. He also highlights Opus 4.8's standout improvement as honesty rather than raw capability — describing it as roughly four times less likely than its predecessor to let flaws in its own code pass unremarked. The post cites Stripe reportedly using Fable 5 to run a codebase-wide migration on a 50-million-line Ruby codebase in a single day, work characterized as equivalent to over two months of team effort. Khaitan's recommended triage flow starts every task at Sonnet 4.6 and climbs the ladder only when that model visibly struggles. As of June 12, 2026, public access to Fable 5 and Mythos 5 is suspended under a US government export-control directive.
Key facts
- 01Anthropic's 2026 model ladder runs: Haiku → Sonnet 4.6 → Opus 4.8 → Fable 5 → Mythos 5
- 02Sonnet 4.6 is priced at $3/$15 per million tokens with a 1M-token context window
- 03Opus 4.8 is priced at $5/$25 per million tokens ($10/$50 in fast mode) and supports dynamic workflows with hundreds of parallel subagents
- 04Fable 5 and Mythos 5 share the same underlying model; the only difference is Mythos 5 has safety classifiers removed and is restricted to vetted partners
- 05Opus 4.8 is described as roughly four times less likely than its predecessor to let flaws in its own code go unremarked
- 06Stripe is cited as reporting that Fable 5 completed a migration of a 50-million-line Ruby codebase in a single day, work estimated at over two months for a team
- 07As of June 12, 2026, public access to Fable 5 and Mythos 5 is suspended under a US government export-control directive
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 20, 2026 · 08:55 UTC. How this works →