Meta Readies Watermelon AI Model to Close Gap With GPT-5.5
Meta is positioning Watermelon as a direct competitor to OpenAI's GPT-5.5, closing a measurable performance gap on coding benchmarks while hinting at further improvements to rival Anthropic's Claude Opus.
Reporting from 1 source: GIGAZINE.
Meta Chief AI Scientist Alexandre Wang revealed that the company's Muse Spark model will be updated under the internal codename Watermelon, which has achieved benchmark scores comparable to OpenAI's GPT-5.5. The update focuses on coding and agent capabilities, with internal tests showing a deliberation mode can improve scores by 8%.
Meta's next AI model, internally codenamed Watermelon, is on track to match OpenAI's GPT-5.5 on key coding benchmarks, according to Meta Chief AI Scientist Alexandre Wang. Wang disclosed the update in a public statement, noting that the Muse Spark model's name will remain unchanged but its internal codename shifts from Avocado to Watermelon. On SWE-Bench Pro, a coding agent evaluation, the previous Avocado version scored 52.5% against GPT-5.5's 58.6%; GPT-5.5 also led on Terminal-Bench 2.0. Meta says Watermelon has closed that gap. The model includes a deliberation mode that uses AI agents to refine responses, which internal tests show boosts scores by 8%. When asked when a model comparable to Anthropic's Claude Opus 4.8-which scores 69.2% on SWE-Bench Pro-would arrive, Wang replied, 'Soon, very soon.' Some information suggests Watermelon's output gains stem partly from increased infrastructure usage rather than pure model improvements.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.