Claude Haiku 5.5 vs Muse Spark 1.3

Where it leads

• Lower input and output prices, $0.10 and $0.50 vs $1.25 and $4.25 per 1M tokens
• Higher FrontierMath Tiers 1-3 (v2) score, 75.1% vs 74.4%
• Cheaper cached input, $0.01 vs $0.15 per 1M tokens
• Batch pricing, $0.05 input and $0.25 output per 1M tokens

Best for

• High-volume classification, extraction and routing under the 100,000-token line
• Cost-sensitive fan-out steps where a frontier model is overkill

Watch out for

• Priced by prompt length, and it is a cliff rather than a taper: a request whose prompt passes 100,000 tokens is billed at $0.50 input and $2.50 output for the whole request, five times the short rate
• Cache reads and cache writes both count toward that 100,000-token measurement, so a large cached prefix can tip a request into the higher band on its own
• It is the only current Claude model whose 1M window is not priced uniformly; every other 4.6-and-later model bills a 900k request at the same rate as a 9k one

Where it leads

• Accepts audio and video input

Best for

• Tool-heavy agent loops where fewer calls per task matters as much as the per-token rate
• Coding agents at volume, where its coding scores and low price line up

Watch out for

• Meta’s own documentation states audio understanding is not fully supported and response quality may be degraded on requests containing audio
• The much cheaper contributor endpoint (about $0.10 / $0.20 per million) is priced that way in exchange for Meta training on your traffic. The standard tier is the rate recorded here.
• Meta’s scorecard has it trailing Claude Opus 5 on every broader agent-workflow benchmark it published, including GDPVal-AA v2, despite winning the coding rows