Here’s what’s shaping AI right now - without the noise:

Zuckerberg closed day one of Meta Connect with a "one more thing." It is Muse Charm, a palm-sized totem with a small screen carrying the Muse avatar, known inside the company as Jolly. You tap a fingerprint sensor in the corner and start talking, with no phone screen to wake and no app to open.

Meta packed the entire real-time voice and avatar stack into something that clips to a keychain. It is not shipping yet. Zuckerberg said the team still has to finalise the component layout, and promised devices in time for the December holidays.

The rest of the keynote went wide. Meta showed VR glasses that weigh one fifth of a Quest 3 and turn any flat surface into a keyboard, plus an always-available agent for its smart glasses line. Read the coverage.

The bigger signal: 👉 Meta is buying distribution in every form factor at once, because the agent that wins is the one closest to hand. Muse already beat ChatGPT's mobile debut with 1.8 million downloads in 12 days. A keychain is just a cheaper phone.

Pioneer, the summit where CX leaders redefine what’s possible, is on October 7th.

Join leaders from Fin, Anthropic, Clay, and Kalshi for an insightful conversation on the state of AI transformation.

You’ll discover how some of the most innovative minds in CX have transformed their organizations, learn how they think about CX, and hear how they're planning for what's next.

Join the conversation in San Francisco, or tune in virtually.

Reserve your spot

Anthropic released Claude Opus 5.5. OpenAI answered with GPT-6 Sol and Luna about 90 minutes later. This is two weeks into an industry that agreed to slow down.

Opus 5.5 takes the top overall spot on Artificial Analysis' Intelligence Index at 58, past Fable 5.1 and GPT-6 Astra at 53. Anthropic says it runs roughly 40% cheaper than Opus 5 by spending fewer tokens, and that it posted the company's best internal alignment score to date. Anthropic also addressed the long-running complaint about "Claudish" writing, saying 5.5 drops jargon and holds closer to your own style rules.

OpenAI's release is priced rather than benchmarked. Sol and Luna post slight gains over their 5.6 counterparts at half the cost, at $0.10 and $0.50 per million for Luna, and $2 and $10 for Sol. Read Anthropic's post or OpenAI's.

The bigger signal: 👉 Sam Altman said pacing does not mean stopping, and this was the proof. If you budgeted for frontier intelligence at last month's prices, you are now overpaying by half.

Anthropic published internal measurements showing Claude "leads" 26% of the company's AI research and development, completing tasks end to end from a prompt while a human supervises. In February that number was under 1%.

Using a scale from Epoch AI, Anthropic puts Claude at "collaborates or above" for more than 90% of its R&D. Full autonomy sits at zero. No task yet reaches the point where Claude fixes a problem and ships it without a human.

The operational detail is the one worth keeping. Anthropic reports roughly 30,000 AI agents running its research and engineering, with every action monitored and 1 in 47,000 blocked. About 6% of R&D compute goes to safety. Read the numbers.

The bigger signal: 👉 Anthropic called for a slowdown last week and then published a dashboard of its own recursive self-improvement. Watch whether OpenAI and SpaceXAI publish theirs.

Opus 5.5 landed on top of the index. Sol and Luna halved their prices. So this week a lot of teams will change one string in a config file and ship it.

That string is the most dangerous line in your codebase.

Why swaps break things silently

A prompt is tuned against a model, whether you meant to tune it or not. Move to a different model, or let a provider roll a version behind the same alias, and your output format drifts, your refusal behaviour changes, and your edge cases start failing without a single error. Nothing alerts. You find out from a customer three weeks later.

Twenty cases beat a test suite

You do not need an evaluation framework. You need a small file of real inputs with known-good outputs, kept in your repo next to the prompt.

Pick 20 cases from your actual traffic:

Keep the expected output next to each one. Run the set against the new model before it reaches a user. Anything that changes shape gets read by a human.

Pin the version, then record what you tested

Never point production at a floating alias. Pin the exact version string, and keep a short table of which prompt was verified against which model.

When a row says untested, that is a blocker and not a note. This table takes 10 minutes to build and it is the thing that tells you which of your 40 prompts actually need rechecking on a release day like this one.

Route by task, not by loyalty

The cheap models got good enough this week that the frontier model is now the wrong default for most calls. Send classification and extraction work to the cheap tier. Keep the expensive model for the two or three steps where it measurably wins on your golden set. Let the numbers pick, rather than habit.

The bigger signal: 👉 Model prices will keep falling and models will keep moving under you. The teams that benefit are the ones who can answer "does the new one still work?" in 10 minutes instead of a week.

A fast AI answer isn't always a coordinated one. When AI resolves B2B issues without looping in the teams responsible for them, confidence outpaces alignment. A new Harvard Business Review Analytic Services briefing paper, sponsored by Front, looks at closing that gap.

Read the Briefing Paper

A few tools quietly worth exploring:

Two weeks into an agreement to slow down, both labs shipped on the same afternoon and cut prices. Anthropic then published that Claude leads a quarter of its own research, and an OpenAI agent turned up inside a government portal. 

Agents got cheaper this week, and they got further inside things. Pacing appears to mean announcing it first.

See you next issue.