Claude Opus 5.5 Is Live on Multos — Fable-Class Work at 40% Lower Cost
Anthropic released Claude Opus 5.5 today and it's immediately available on Multos Elite. It matches Claude Fable 5.1 on most tasks, costs 40% less to run than Opus 5, and is 30% faster. On Terminal-Bench 4.0 it scores 66.4% — higher than GPT-6 Astra, higher than Opus 5, higher than Fable 5.1. Here's what that actually means for building apps and games on Multos.
Anthropic dropped Claude Opus 5.5 this morning, September 22, 2026, and it's available on Multos Elite starting today. This is the first model in a new Claude 5.5 family — Sonnet 5.5 and Haiku 5.5 are coming in the next few weeks — and Anthropic's pitch is sharp: frontier-level work, fewer tokens per task, and a price cut. $4 per million input tokens versus $5 for Opus 5. $20 per million output versus $25. Cache reads at $0.20 versus $0.50 — a 60% drop on the number that matters most for agentic and coding workloads. On Multos, where Auto Router makes heavy use of cached context across turns, that cache reduction isn't a small detail. It's most of your bill.
What Anthropic Actually Claims
The headline number is Terminal-Bench 4.0: 66.4%. GPT-6 Astra sits at 57.9% on that benchmark. Opus 5 is at 52.3%. Fable 5.1 is at 55.8%. Opus 5.5 is ahead of all of them. FrontierCode v1.1 Main: 54.4%. GPT-6 Astra: 53.3%. Opus 5.5 beats Astra here too, and does it at about one-fifth the cost per task. CursorBench 4.0: 57.8%. GPT-5.6 Sol: 41.7%. That's an 11-point gap over Sol, for roughly a third of the cost. OSWorld 2.0 (computer use): 81.8%. Fable 5.1: 80.7%. Opus 5: 74.4%. Humanity's Last Exam (with tools): 67.7%. Fable 5.1: 65.6%. These are the hardest reasoning benchmarks in circulation, and Opus 5.5 is leading them. Knowledge work via GDPval-AA v2.1: 1846 Elo. Fable 5.1: 1735. Opus 5: 1708. GPT-6 Astra: 1542. The benchmark story is unambiguous — Opus 5.5 leads on agentic coding, computer use, and knowledge work. Where Anthropic is careful to be honest is the efficiency angle: it's not just that prices dropped, it's that the model uses fewer tokens per task at the same or better quality. An audit-and-fix job on a 200,000-line codebase that took Opus 5 over 20 hours took Opus 5.5 under three hours, using 2.5x fewer tokens.
What This Means for Building on Multos
If you're building apps on Multos and you've been reaching for Opus 5 for the hard jobs — complex full-stack generation, architectural decisions that touch multiple files, multi-step debugging that requires the model to hold a lot of state — Opus 5.5 is the upgrade you didn't have to wait for. Same tier, same access, 40% less spend per session. For game development specifically: Opus 5.5's computer use score (81.8% on OSWorld 2.0) means it can navigate interfaces, test layouts, and reason about what a user sees in a screenshot without you describing it in text. You can share a frame from your Phaser game, describe the bug, and the model understands the visual context. That's not a gimmick — it cuts a feedback loop that used to require you as the translator between what the model generated and what actually rendered. For app builders: the multi-agent architecture gains are practical. Anthropic's testing showed Opus 5.5 delegating to subagents far more reliably, self-verifying work in ways that don't require another human prompt to get unstuck, and — critically — generating output that reads like a competent colleague wrote it rather than a model straining to be helpful. One real tester's quote from Anthropic's launch post: 'It writes the way I do.' That's a specific and useful claim for anyone whose product puts Claude's output in front of end users.
Real Production Numbers
Anthropic's enterprise partners published their own data alongside the launch. Some of these are worth reading as building blocks for what you might expect: Optiver (trading firm): Claude Opus 5.5 matched Opus 5's quality in about half the turns, half the time, and half the output tokens on their agentic coding tasks — 40-50% cost reduction on those workloads. Spotify: 'Clear improvement in token efficiency... complete the same tasks both cheaper and faster.' Stripe: A multi-day rebase of 40 stacked pull requests. One Opus 5.5 session directed a dozen sub-sessions, laid out every conflict, and all 40 passed CI the next afternoon. GitHub: Opus 5.5 solved more terminal tasks than Opus 5 in less than half the steps in VS Code. These are enterprise shops with custom harnesses, but the pattern is consistent: fewer steps, less rework, cheaper per meaningful unit of output. For Multos specifically, where sessions run across Auto Router, context compression, and multi-agent subagent pipelines, fewer steps per task directly reduces how fast your token pool drains.
Building Full-Stack Apps
Here's a concrete workflow for building a full-stack web app on Multos with Opus 5.5. Start by uploading your entire repository or pasting your architecture notes. Opus 5.5's 200K context window means your whole project fits. Ask it to audit the existing structure before writing a single line of new code. It will catch problems upstream — Walleye Capital noted it found an indexing bug in their own instructions, corrected for it, and noted it would have cost them points with the grader. That kind of catch happens on your app's schema before it becomes a production bug. For the generation pass, give it a full spec. Not a bullet list — a proper spec with input/output contracts, edge cases, and the stack you're deploying to. Opus 5.5's FrontierCode score (54.4%) reflects the kind of tasks where the model has to decide whether to open a new file, modify an existing one, or create a test first. It handles that decision tree better than any model before it. For games and interactive apps, the visual loop matters. Screenshot the running app, paste it, ask what's wrong. The computer use capability that scored 81.8% on OSWorld doesn't require a special setup — it's the same model in the same chat. You describe the behaviour you're seeing and it can reason about the visual state, not just the code.
The Safety Note
One thing worth knowing if you're building production systems: Opus 5.5 has stronger prompt injection defences than Opus 5 — it matches or beats Opus 5 in every injection setting Anthropic tested. It also has a classifier that screens every action before it runs, and code review that catches vulnerabilities before they merge. For agentic applications where Opus 5.5 is running unattended — overnight jobs, batch processing pipelines, CI-integrated code reviews — that's not a soft benefit. It means the model is less likely to be hijacked by content it encounters while doing the work. The 85% reduction in containment boundary crossing attempts compared to Opus 5 is the number Anthropic leads with, and it's independently meaningful for anyone shipping autonomous agents to production.
Where to Find It
Claude Opus 5.5 is available on Multos Elite today. Select it from the model dropdown under the Anthropic group, or let Auto Router route to it for complex, multi-step tasks where it's the right cost-performance fit. For Pro plan users who want to try it: it's Elite-tier on Multos, which means upgrading to Elite unlocks it alongside GPT-6 Astra, GPT-6 Luna, and Claude Opus 5. The model runs through Multos's managed Anthropic connection — no API key required, no setup needed.
Ready to Get Started?
Claude Opus 5.5 is live on Multos Elite. The strongest agentic coding model Anthropic has ever shipped, at 40% lower cost than Opus 5. Select it from the model dropdown or let Auto Router pick it for your hardest tasks.
Start Building FreeRelated Articles
GPT-6 Sol and GPT-6 Luna Are Live — OpenAI Halves the Price of Its Best Models
OpenAI released GPT-6 Sol and GPT-6 Luna today, cutting prices by 50% versus GPT-5.6. Sol drops from $4/$20 to $2/$10 per million tokens. Luna drops from $0.20/$1.20 to $0.10/$0.50. Intelligence Index scores hold steady, hallucinations drop significantly, and both models are live on Multos now — Sol on Pro, Luna on Lite.
Read more →MiMo V2.6 Flash and Pro Are Now on Multos — Frontier-Grade Agents at Flash-Tier Pricing
Xiaomi released MiMo V2.6 Pro and Flash on September 21, 2026. Pro scored 46 on the Artificial Analysis Intelligence Index — the highest any open-weight model has ever scored, at $0.13 per task. Flash costs $0.14/$0.28 per million tokens and trails Pro by only 4 points on DeepSWE. Both are live on Multos Lite via DeepInfra.
Read more →GLM-5.3 Flash: The Model That Was Hiding in Plain Sight
For six days, one of the most capable models on OpenCode and OpenRouter had no name. It was just called Ox Alpha. Then Z.ai revealed it: GLM-5.3 Flash, a 320B MoE model with a 1M token context, native multimodal input, MIT-licensed weights — and a price tag that makes you read it twice.
Read more →