Home Posts Not All Model Upgrades Are Upgrades [Evals]
Engineering

Not All Model Upgrades Are Upgrades [Evals]

Dillip Chowdary
Dillip Chowdary
July 21, 2026 · 7 min read · Source: Microsoft Developer Blog

Bottom Line

A newer model with a cheaper rate card can still be more expensive. Microsoft measured Claude Sonnet 5 vs 4.6 on agent tasks and saw median token multipliers up to ~12× on architecture work — enough to erase a 33% per-token discount.

Key Takeaways

  • Sonnet 5 rate card in the study: $2/$10 vs Sonnet 4.6 $3/$15 (input/output per 1M) — Sonnet 5 cheaper per token.
  • Architecture tasks: ~12× more tokens at the median for Sonnet 5; some scenarios much higher.
  • Code upgrade tasks: ~10× token gap; example cost $2.01 vs $0.55 per run in the write-up.
  • Evaluate upgrades with fixed harness + fixed tasks; never with “feels smarter.”
  • Gate production model flips on cost-per-success, not benchmark screenshots.

Microsoft’s Agent Experience team published “Not all model upgrades are upgrades” with hard numbers from Claude Sonnet 4.6 vs Sonnet 5 agent runs.

Rate card vs bill

MetricSonnet 4.6Sonnet 5
Input / 1M$3.00$2.00
Cached input$0.30$0.20
Output / 1M$15.00$10.00
Architecture tokens (median)baseline~12× higher
Example code-upgrade $ / run$0.55$2.01

Method (copy this)

  • Fixed product surface (GitHub Copilot Chat in VS Code)
  • Two task classes: architecture/design grounded in docs, and SPFx project upgrades
  • Many repeats per scenario to capture variance
  • Compare tokens and dollars, not only pass rates

When to choose each

  • Choose the newer model when: quality or autonomy gains outweigh measured $/success and you have budget headroom.
  • Stay on the older model when: token multipliers destroy unit economics or the new model thrashes tools without better outcomes.

Harness controls that reduce upgrade risk

  1. Cap max tool calls and max dollars per task
  2. Force plan → execute → verify phases
  3. Cache prompts and shared context aggressively
  4. Run dual-model shadow traffic for a week before flipping defaults
Primary source: Microsoft Developer Blog → Verify claims against the original before changing production systems.

More from Tech Bytes