Opus 5.5 and GPT-6 shipped on the same afternoon, both cheaper, and the fine print matters more than the headline
Wednesday, September 23, 2026·8 min read·4 stories
Two frontier labs cut prices on the same Tuesday, and neither mentioned the other. Opus 5.5 is 20% cheaper per token and claims 40% cheaper per task. GPT-6 Sol and Luna are about half the price of the 5.6 series and, per an independent index, about as smart. The real savings for most teams are not in either headline. They are in two quieter releases from the same 24 hours: OpenAI's rebuilt prompt caching, and an open-source harness from AWS that attacks the other half of your bill, the tokens your agent wastes on its own context. Then Meta's new agent reminded everyone what happens when a desktop agent trusts every process on the machine.
No sponsored or affiliate links in this digest — the links below are sources only.
Story I
Opus 5.5 vs. GPT-6 Sol and Luna: the price sheet, and the output-token catch
Anthropic's Claude Opus 5.5 (claude-opus-5-5) went live September 22 at $4 per million input tokens and $20 output, down from $5 and $25. Cache reads dropped from $0.50 to $0.20 and cache writes from $6.25 to $5. Anthropic's 40% figure is per task, not per token: the 20% price cut plus fewer tokens spent and output that arrives about 30% faster. On benchmarks it beats Fable 5.1, the company's bigger model, on Terminal-Bench 4.0 (66.4% vs 55.8%) and FrontierCode (54.4% vs 50.3%). It is on the Anthropic API plus AWS, Google Cloud and Azure. Anthropic also admitted the "Claudish" writing complaints and says the model now leads with the important part and follows style instructions more closely.
OpenAI's GPT-6 Sol landed the same afternoon at $2 input and $10 output, down from $4 and $20. Luna, the high-volume tier, is $0.10 and $0.50. The catch is on the scoreboard: Artificial Analysis moved Sol from 47 to 48 on its Intelligence Index and left Luna at 37, and Sol still trails the older Opus 5 on FrontierCode 1.1 (49.3% vs 53.4%). This release is a price cut wearing a version number.
Now the fine print on Opus. At maximum effort it spends around 119,000 output tokens per task, well above its rivals, which eats into the per-token savings. Simon Willison hit the 128,000-token output ceiling mid-reasoning on a simple SVG prompt, and each failed attempt cost him $2.56 and 20 minutes. He has still switched his defaults to Sol and Opus 5.5.
For builders
Don't migrate on list price. Pull 50 real requests from yesterday's logs and replay them against claude-opus-5-5 at your current effort setting and at one level lower, and against gpt-6-sol, then compare total output tokens and cost per completed task, not per token. If you run Opus at maximum effort by default, cap it: set an explicit output-token limit and only escalate effort on retries. And if you cache heavily on Anthropic, recompute your bill, because the 60% cut on cache reads may be worth more to you than the headline 20%.
Story II
OpenAI rebuilt prompt caching for GPT-6, and it now tells you why you missed
The quieter OpenAI post from September 22 matters more for anyone running agents. Cache discounts of up to 90% on cached input now apply to eligible shared prefixes reused within a 30-minute window, and the new caching system raises hit rates by default. There is a Prompt Caching Dashboard in the usage console that shows hit rate over time and cached versus uncached input.
The useful piece is the diagnostics tool. Compare a request against a recent one and it returns a structured reason for the miss, like "reason": "tools_changed" with the count of tokens that would have been reusable. OpenAI also added explicit cache breakpoints, lets you change reasoning effort mid-conversation with a configuration_update without breaking cache, and supports prewarming shared instructions and tool definitions at startup.
The customer numbers are vendor-picked but specific. Manus says its cache hit rate went from roughly 85% to consistently above 90% in under a week. Wordsmith moved session agents to explicit breakpoints and saw hit rates go from 83% to 91%, cache writes fall by about two thirds, and inference costs drop 36%.
For builders
The most common cache-killer in agent code is editing the tool list between turns. OpenAI's own guidance: keep tool definitions, schemas and ordering stable, and use allowed_tools or tool_choice: none to restrict what is callable instead of deleting definitions. Append new instructions as a new developer message at the end rather than rewriting the system prompt. Open the dashboard today, find your lowest-hit route, run one of its requests through the diagnostics tool, and fix whatever reason it names first.
Story III
AWS open-sourced Strands Harness and says it undercuts Claude Code on cost
The AWS Strands Agents team released Strands Harness, an Apache 2.0 agent harness for Python (pip install strands-harness) and TypeScript (npm install @strands-agents/harness). It is a general-purpose agent loop, not a coding-only tool, and works with Bedrock, Anthropic, OpenAI, Google, Ollama and LiteLLM. The repo is strands-agents/harness-sdk on GitHub.
The pitch is token discipline. Tool results over about 1,500 tokens get truncated, bulky results are offloaded to files, context is compacted at 85% usage, open-ended subtasks go to a helper subagent, and reused request parts are cached. AWS reports 28% lower token cost on average across six benchmarks (ALFWorld, ContextBench, GAIA, WebShop, τ²-bench and Terminal-Bench 2.1) against other harnesses running the same Claude or GPT models. On Terminal-Bench 2.1 it claims 77% lower cost than Claude Code with 7.9 points higher accuracy.
Treat those as AWS's numbers on AWS's chosen benchmarks. Still, the direction lines up with the caching story above: the model is only half the bill, and the loop around it decides how many tokens you burn.
For builders
You don't have to switch harnesses to take the idea. Look at your agent's traces for the three largest tool results per run. If any tool routinely returns 10,000-token blobs, such as full file reads, raw HTML or unfiltered JSON, truncate them to a summary plus a file path the agent can open on demand. That one change is the core of what Strands is selling. If you want the full thing, install it in a scratch repo and run it on one of your own existing tasks with the same model before trusting the 28%.
Story IV
Meta's Muse agent let any local process steal its auth token. Build yours so it can't.
Muse is Meta's new macOS agent. It books appointments, fills forms, makes purchases, and connects to WhatsApp, email, calendars and social accounts, with OS permissions to write to disk, use the mic and camera, and track location. Patrick Wardle found that any locally run app or terminal command, whatever its macOS permissions, could change a long list of undocumented Muse settings. One of them sets the endpoint for voice transcription. Point it at an attacker's server and that server receives the account token, and can proxy traffic to Meta while injecting prompts like exporting all WhatsApp messages.
Wardle's proof of concept wrote files to disk and took photos with no visible sign. A ClickFix-style lure, tricking the user into pasting a command, is enough to deliver it. Meta shipped a hotfix more than 12 hours after the disclosure went live and called it "not a remote exploit." Hours before the disclosure, Amazon had started blocking Muse as an "unauthorized AI agent" that violates its Conditions of Use.
For builders
If you ship a desktop or CLI agent, check three things this week. First, is any config that controls where data or credentials go, such as endpoints, proxies or base URLs, writable by any process running as the user? If so, sign it, pin it, or require a user confirmation to change it. Second, is the auth token stored in the OS keychain with per-app access, or sitting in a plaintext file? Third, does the agent refuse to send its token to a host outside an allowlist? Muse failed the first and the third. Also read the Amazon line closely: if your agent shops or books on third-party sites, those sites can and will block it.