Mid-Tier Models Are the New Default

Updated 2026-08-24: as published, this post priced Claude Sonnet 5 at $3.00/$15.00 and closed with an "introductory window" expiring August 31, 2026. Anthropic cancelled the increase that window pointed at, and our registry has carried the real $2.00/$10.00 rate since 2026-07-21. The Sonnet 5 figures, the routed-workload arithmetic and the closing section are corrected below; the ranking and confidence data is as published.

The mid tier has caught up

Twelve months ago, mid-tier models were a compromise. In July 2026, they lead the rankings in four of eight task types.

Here are the top-scoring models per task type, with their tier:

• -Code review: Claude Sonnet 5, 96/100 (mid tier)
• -Summarization: GPT 5.6 Terra, 92/100 (mid tier)
• -QA: GPT 5.6 Terra, 92/100 (mid tier)
• -Extraction: GPT 5.6 Terra, 90/100 (mid tier)
• -Classification: Gemini 3.5 Flash, 92/100 (fast tier)
• -Reasoning: Claude Fable 5, 98/100 (flagship)
• -Code generation: Claude Fable 5, 96/100 (flagship)
• -Creative: Claude Fable 5, 95/100 (flagship)

Flagships lead only on reasoning, code generation, and creative tasks. Mid-tier and fast-tier models own the rest.

The price gap is 50-60%

Mid-tier models cost $2.00 to $3.00/MTok input versus $5.00 to $10.00 for flagships. Here is the savings breakdown:

• -Gemini 3.1 Pro ($2.00) saves 60% versus Opus ($5.00) on input
• -GPT 5.6 Terra ($2.50) saves 50% versus GPT 5.6 Sol ($5.00) on input
• -Claude Sonnet 5 ($2.00) saves 60% versus Opus ($5.00) on input

Output savings follow the same pattern, ranging from 40% to 70% per million tokens.

Real-world savings for a mixed workload

Consider a team running 50,000 API calls per day across five task types. Average call uses 3,000 input and 2,000 output tokens.

All flagship (Opus 4-8):

• -Input: 50,000 x 3,000 / 1,000,000 x $5.00 = $750/day
• -Output: 50,000 x 2,000 / 1,000,000 x $25.00 = $2,500/day
• -Total: $3,250/day, $97,500/month

Task-routed mid-tier (Sonnet 5 for code, Terra for QA/extraction):

• -Input: 50,000 x 3,000 / 1,000,000 x $2.25 (avg) = $337.50/day
• -Output: 50,000 x 2,000 / 1,000,000 x $12.50 (avg) = $1,250/day
• -Total: $1,587.50/day, $47,625/month

Monthly savings: $49,875 by routing to mid-tier models. Quality stays within 5 confidence points.

When flagships still make sense

Flagships earn their price on three workload types:

• -Complex reasoning chains where 97+ confidence scores are required for accuracy
• -Novel code generation on unfamiliar frameworks where Fable 5 scores 96/100
• -High-stakes creative writing where the 8-point gap between tiers affects output quality

For everything else, mid-tier models deliver comparable quality at a fraction of the cost.

Sonnet 5 at $2.00/$10.00

Claude Sonnet 5 runs at $2.00/$10.00. Of the mid-tier models compared here, that makes it the cheapest option for code-heavy workloads, and it is the top-ranked model for code review.

Anthropic announced that rate as introductory pricing through August 31, 2026, then cancelled the increase to $3.00/$15.00 that was meant to follow. There is no window to evaluate inside: the price is the price, and a team that switches keeps the saving.

Run your task-type breakdown at /teardown, or start tracking with a free dashboard.