PropagandAI

GPT-6 Sol and Claude Opus 5.5 land 90 minutes apart, and prices fall

OpenAI and Anthropic shipped cheaper frontier models on the same day, OpenAI opened training to outside testers, and two security stories showed where AI risk is moving.

By ·

Abstract illustration of two glowing towers rising side by side as price-tag shapes fall between them, with a padlock and circuit lines below.
Illustration generated with AI

Tuesday brought two frontier model launches within about an hour and a half of each other, and both led with lower prices rather than just higher scores. Around them: a new promise from OpenAI about outside safety testing, a local flaw in Meta's fast-growing Muse assistant, and an open-source tool from Cisco for tracking malware that calls AI models.

OpenAI ships GPT-6 Sol and Luna at half the old price

OpenAI released GPT-6 Sol and Luna, two new members of its GPT-6 generation. According to TechCrunch, Sol is aimed at harder work such as coding, while Luna targets high-volume jobs like summarizing documents and pulling out information. Both cost half as much as the models they replace in the 5.6 series, which OpenAI credits to caching and inference improvements. OpenAI claims Sol makes roughly half as many mistakes as its predecessor. The models are available in the API, ChatGPT Work and Codex for paid users, with Luna also coming to free users.

Why it matters: OpenAI is competing on cost per task, not only on capability. For developers, the same budget now buys considerably more work.

Anthropic answers with Claude Opus 5.5

About 90 minutes later, Anthropic launched Claude Opus 5.5, the first model in its Claude 5.5 family. Anthropic says it performs close to its larger Fable 5.1 on most work while cutting prices to $4 per million input tokens and $20 per million output tokens, down from $5 and $25. The company expects about 40% savings on typical workloads and says output is about 30% faster than Opus 5. It reports 66.4% on Terminal-Bench 4.0 for agentic coding. Anthropic says METR and another outside group evaluated the model before release, and that most high-risk cybersecurity requests are routed to an older model as a safeguard. It is available on the Claude Platform, AWS, Google Cloud and Microsoft Azure.

Why it matters: With both labs cutting prices on the same day, the frontier race now looks like a price war as much as a benchmark contest.

OpenAI will let outside groups test models during training

OpenAI also said it will open technical safety assessments to outside organizations earlier in a model's life: during training and evaluation, not only just before launch. The Next Web reports that OpenAI is in talks with METR and Redwood Research, but no formal partnerships or start date were announced. The outlet also notes the written plan is vaguer than what Sam Altman described ten days earlier, when he spoke of evaluators getting desks and laptops with the right to publish. Bloomberg first reported the change.

Why it matters: Testing before a model is finished can catch problems when they are cheapest to fix, but the value depends on access terms OpenAI has not yet published.

A local flaw in Meta's Muse assistant

Meta's Muse assistant, which TechCrunch says recently reached No. 1 on the U.S. App Store, has an unpatched local vulnerability, first reported by Ars Technica. A proof of concept on GitHub shows that an undocumented setting can be changed by an ordinary local process to send Muse's dictation traffic to an attacker's server. That could expose dictated audio and prompts, allow injected prompts, or leak credentials. The attack requires code already running as the user. Separately, Meta told TechCrunch that Muse was heavily inspired by the open-source agent OpenClaw, though built from scratch.

Why it matters: AI agents that get broad access to a computer become attractive targets themselves. A small settings bug matters more when the app holding it can do a lot.

Cisco open-sources CAIRN to track AI-powered malware

Cisco Talos released CAIRN, an open-source toolkit for finding and classifying malware that uses AI models. Instead of running samples, it looks at their metadata, such as API keys, prompt templates and model provider endpoints. Talos says its research found malware moving within about a year from optional AI features to fully autonomous setups that coordinate several models, and that AI-specific evasion tricks are spreading between malware families. The code is on GitHub.

Why it matters: Defenders now have a shared, free way to spot malware that leans on commercial AI services, using traces those services leave behind.

Quick answers

What did OpenAI and Anthropic release on September 22, 2026?

OpenAI released GPT-6 Sol and Luna at half the price of the 5.6-series models they replace, and Anthropic released Claude Opus 5.5 at $4/$20 per million tokens, down from $5/$25 (TechCrunch, Anthropic).

Will outside groups test OpenAI's models before launch?

OpenAI says outside organizations will be able to run safety assessments during training and evaluation, and it is in talks with METR and Redwood Research, but no partners or start date have been announced (The Next Web).

What is the Meta Muse vulnerability?

A proof of concept shows an ordinary local process can change a hidden Muse setting to redirect dictation traffic to an attacker, though the attacker must already be running code on the machine (GitHub).

Sources