Claude Sonnet 5.5 Beats Opus 5.5 On Terminal-Bench At Half The Price And Runs 30% Faster!

Today's AI news: Anthropic released Claude Sonnet 5.5 at Sonnet 5's price, more than 30% faster, scoring 70.6% on Terminal-Bench 4.0 to beat Opus 5.5's 66.4%, and 80.1% on OSWorld. Manus launched Manus 2.0 on its Cascade architecture, 32% cheaper to run, plus Cue, a personal agent app where each agent gets its own email, phone number, wallet and computer. ElevenLabs released Eleven v4, number one on Artificial Analysis's voice leaderboard, with 90+ languages and a 100ms Turbo variant. And Starship reached orbit for the first time, deploying 26 Starlink V3 satellites. Chapters: 0:00 Today's AI News 0:25 Claude Sonnet 5.5 1:17 Manus 2.0 1:59 ElevenLabs v4 2:32 Starlink V3 Satellites Blend Roundup 2026-09-28 https://x.com/claudeai/status/2104633115620823187 https://x.com/claudeai/status/2104633119412523294 https://x.com/ManusAI/status/2104592659142869352 https://x.com/CueAgents/status/2104589146153181192 https://x.com/ManusAI/status/2104628185501712789 https://x.com/ElevenLabs/status/2104572127617994917 https://x.com/niccruzpatane/status/2104568103174566043 Anthropic has released Claude Sonnet 5.5. On some coding benchmarks it beats Opus 5.5 at half the price. Manus has launched version 2.0 plus an app that gives your agents their own phone numbers, ElevenLabs has a new voice model at the top of the leaderboard, and Starship has finally reached orbit with the first Starlink V3 satellites. Here's today's AI news. Anthropic has released Claude Sonnet 5.5, the second model in the Claude 5.5 family, and the gap to Opus has almost closed. It keeps Sonnet 5's price, runs more than 30 percent faster, and Anthropic says it costs up to 30 percent less for most work. The jumps over Sonnet 5 are huge. On Terminal-Bench 4.0 it scores 70.6 percent against Sonnet 5's 10.3, and it actually beats Opus 5.5, which gets 66.4. On OSWorld computer use it hits 80.1 percent, within two points of Opus, and it's level with Opus on the knowledge-work tests. Opus still leads on most other tests, but only by a few points. Anthropic pitches it as the faster, cheaper complement to Opus, strongest at well-scoped everyday tasks - including fixing bugs and creating polished documents, slides, and spreadsheets. Manus has launched Manus 2.0, built on a new agent architecture called Cascade. Manus says it uses 23 percent fewer tokens, finishes tasks 28 percent faster and costs 32 percent less to run than before. Workflows can now trigger on emails, Slack messages or calendar changes, and a new Studio adds tools for video editing, game development, documents and code. The bigger idea is Cue, a separate app for personal agents. Each agent gets its own email, phone number, wallet and computer, so it can make and take calls, send texts, and pay for things within a budget you set. Cue is invite-only in early access for now. ElevenLabs has released Eleven v4, its most emotive voice model yet, and it's gone straight to number one on Artificial Analysis's voice leaderboard. ElevenLabs says listeners preferred it around 75 percent of the time in blind tests. It supports more than 90 languages, including Cantonese and Mongolian. It keeps a speaker's voice consistent over long recordings, and can clone a voice from ten seconds of audio. There's also Eleven v4 Turbo, built for live voice agents, with around 100 milliseconds of latency. SpaceX's Starship has reached orbit for the first time, and it brought cargo. On Flight 14 it deployed 26 Starlink V3 satellites one by one, and SpaceX has made contact with all of them. Elon Musk says they're operating nominally. Each V3 handles about ten times the data of the V2 Minis that Falcon 9 carries, and they're too big for Falcon 9 to launch at all. At full scale, one Starship will carry 60, the network capacity of about 20 Falcon 9 launches. It wasn't flawless: an engine cut out early, so SpaceX brought the ship home ahead of schedule.