読書

Anthropic and OpenAI released new model tiers about 90 minutes apart on September 22, and both cut prices sharply: frontier-level intelligence keeps falling in price per token.

Anthropic's Claude Opus 5.5 is priced at $4 input and $20 output per million tokens, with cache reads at $0.20; the company says typical workloads cost 40% less than Opus 5, and cache reads 60% less. OpenAI's GPT-6 Sol is 50% below GPT-5.6 pricing at $2/$10, and Luna at $0.10/$0.50 per million tokens.

All performance-leadership claims are vendor-reported: Anthropic itself concedes that benchmark margins have become a less reliable guide at this capability level, and OpenAI's AutomationBench comparisons use its own framing. Treat the price cuts as verified facts and the rankings as claims awaiting independent checks.

Anthropic shipped Opus 5.5 and made it the default model, with list prices down 20% — but real costs at high effort barely moved.

Launched September 22, the model is claimed to match Fable 5.1 on most tasks and run 40% cheaper than Opus 5. Input/output pricing fell from $5/$25 to $4/$20 per million tokens, and it is now the default in Claude Code and the Claude app.

Artificial Analysis, however, measured that higher token usage at max effort would raise per-task cost about 80%; after the price cut and cheaper caching it lands at $5.98 per Intelligence Index task, essentially flat versus Opus 5's $5.86. The 40% saving holds only at default effort.

Performance figures are vendor and partner numbers, and the unconfirmed claim that the model is smaller remains speculation; test cost on your own workload and effort setting before switching.

Anthropic released Claude Opus 5.5 on September 22, cutting input to $4 per million tokens and output to $20, with cache reads down 60% to 20 cents; the faster serving mode runs $8 and $40.

The model is live on the Claude Developer Platform and on AWS, Google Cloud and Azure, with Sonnet 5.5 and Haiku 5.5 due in coming weeks. OpenAI released GPT-6 Sol and Luna at half price the same day, so the frontier price war moved at both labs at once.

Benchmark figures such as 66.4% on Terminal-Bench 4.0 are Anthropic's own, with no third-party verification yet; for teams running long-session agents, the cache price cut is the most certain cost change available today.

Anthropic released Opus 5.5 on September 22, cutting output token pricing to $20 per million tokens from $25 for the previous model, a 20% drop.

The company calls it the strongest-performing model it has tested and says it outpaces its own larger Fable model on many benchmarks. Those results are all vendor-run; METR and other outside groups did pre-release safety evaluation only, and no independent performance verification exists yet.

Anthropic says the model is comparable to Mythos in biology and cybersecurity capabilities, so it carries the same usage safeguards as Fable. Sonnet 5.5 and Haiku 5.5 are promised in the coming weeks; whether the price cut extends to mid-tier models is what buyers should watch.

OpenAI released GPT-6 Sol and Luna on September 22, pricing the API at half the cost of the 5.6-series equivalents, which the company attributes to improvements in caching and inference.

Sol targets complex tasks like coding; Luna handles high-volume clerical work such as summarizing and extraction. Both are live in ChatGPT Work, Codex and the API, and Luna will also reach Free and Go users.

The reliability claim deserves a discount: the reported halving of mistakes comes from OpenAI's internal evaluation based on user-flagged conversations, not third-party testing. Anthropic shipped Opus 5.5 just 90 minutes earlier, and the pricing race between the two is clearly deliberate.

OpenAI released GPT-6 Sol and Luna with API prices roughly half the previous generation's; Luna costs $0.10 per million input tokens and $0.50 per million output tokens.

Per ifanr's report, Luna's standard-request pricing now sits in the range of DeepSeek V4.1 Flash's cache-miss rates (1 yuan input, 4 yuan output per million tokens), though DeepSeek's off-peak cache-hit input at 0.02 yuan remains far below Luna.

Performance figures are OpenAI's own claims; third-party Artificial Analysis estimates both models' intelligence is roughly flat versus GPT-5.6, with the main change being cost.

Epoch AI estimates that over the past three years, the cost of reaching a given level of AI performance has fallen about 47% per quarter on average — roughly 13-fold a year.

The report's example: in January 2025, o3 needed about $0.30 per question to score 75% on GPQA Diamond; in mid-2026, GPT-5.6 Luna hit the same score for $0.0004 — a roughly 725-fold drop in under 18 months. Math benchmarks fell fastest (50–52% per quarter), game-based puzzles slower (39–43%).

This is the authors' own analysis, and the report names its limits: benchmarks may be gamed, and it assumes users always pick the cheapest capable model, so the numbers should not be read as exact. Even at these prices, the report notes total spending could stay high as usage expands.

Alibaba's Qwen team released Qwen-Audio-3.1, a lineup of five audio models, and cut its voice API prices: TTS drops about 70 percent, Realtime roughly 85 percent, and ASR up to 95 percent.

The series covers speech recognition, text-to-speech and real-time interaction. Qwen says the ASR model improves multilingual and dialect recognition and cleans up filler words, TTS-Next generates voice and sound effects in a single diffusion pass, and the real-time model supports simultaneous speaking and listening with instant interruption.

All capability claims are Qwen's own; per The Decoder, they come from the official blog and an X announcement, with no independent evaluation yet. Cost assumptions for voice applications can be repriced at the new rates.

Meta's Muse assistant, launched only weeks ago, shipped with a zero-day: any local app or terminal command could rewrite the transcription server address and capture the account token, taking full control of the agent. Meta released a hotfix roughly 12 hours after disclosure.

The discoverer, macOS security expert Patrick Wardle, says attackers need not write full malware — the agent's own privileges suffice to write files or snap pictures with little or no indication. The root causes are design choices: cloud-based transcription, and any local process being able to control undocumented settings. The patch closes the hole; those decisions stand.

A separate fact: about 12 hours before disclosure, Amazon began blocking Muse from shopping on its site, calling it an unauthorized AI agent that violates its Conditions of Use, and asked Meta to remove Amazon from the experience. The exploit details are Wardle's own proof-of-concept, not independently reproduced; he plans to present them at a security conference in November.

Internal Meta posts seen by 404 Media confirm that the Muse feature, which calls businesses on a user's behalf, includes a human agent layer during testing, with some calls placed by trained human agents.

On September 16 Meta executives publicly promoted Muse's outbound calling. The internal posts state that Muse can hand a request to a trained human agent who places the call. Employees questioned users unknowingly handing call requests to people, and one warned of severe negative coverage.

A Meta spokesperson said this is internal dogfooding and that proper disclosures will come before any public release. How often calls are routed to humans, and under what conditions, remains undisclosed.

Meta's personal AI agent Muse drew more than 500,000 users in the week after its September 8 launch, including over 250,000 daily active users, according to internal data reported by The Information. The app has reached number one in Apple's App Store.

Nat Friedman, head of product at Meta Superintelligence Labs, wrote on X that Muse is "definitely heavily inspired as a product by OpenClaw" but was built from scratch; users had noticed nearly identical file names and contents, and Friedman replied that OpenClaw's creator "got those things exactly right." The user figures are Meta internal data relayed by media, not independently verified.

According to The Information, OpenAI has discussed building its own personal assistant in response. The real test is whether users will hand personal accounts and sensitive data to agents like this.

Two weeks after launch, Meta shares are up more than 20% cumulatively, adding over $200 billion in market value and hitting a seven-month high.

Per a Reuters report (relayed via IT Home), Apptopia data shows Muse reached 2.8 million downloads in its first 12 days, available only in the US and Canada; on a like-for-like basis with ChatGPT's first 12 days, Muse logged 1.8 million downloads versus ChatGPT's 1.3 million. LSEG data shows 57 of 64 covering brokers rate Meta a buy or better.

Jefferies' $10.8 billion annualized revenue figure is a scenario, not a forecast: it requires 1 billion users by end-2027 with at least 3% converting to paid. Downloads are not retention, and paid conversion is unproven.