Anthropic's Opus 5.5 lands 40% cheaper, OpenAI halves GPT-6 API prices 90 minutes later, and 3 more AI moves
Five verified developments from September 22 to 23, 2026: Anthropic's cheaper Opus 5.5 and OpenAI's half-price GPT-6 Sol and Luna released 90 minutes apart, Trump's 'super intelligence' rename at the UN, Beijing's probe of DeepSeek and Moonshot, and Snorkel AI's $350 million round.
In brief
Across the 24 hours from September 22, 2026 at 12:00 UTC through September 23, 2026 at 12:00 UTC, two leading labs put cheaper models on sale 90 minutes apart, two governments went on the record about who may do what with AI, and $350 million went to the company that supplies the labs with training data:
- Models: Anthropic released Claude Opus 5.5, which it says matches its larger Fable 5.1 model for most work and costs 40% less to run than the Opus it replaces, at $20 per million output tokens against $25. 12
- Models: OpenAI released GPT-6 Sol and GPT-6 Luna 90 minutes after that, halving its API prices to $2 and $10 per million tokens for Sol and $0.10 and $0.50 for Luna. 34
- Regulation: Trump told the UN General Assembly that US government documents will call the technology "super intelligence" from now on, that the Justice Department will watch the industry closely, and that he will not "stifle growth of something that will be bigger than the Industrial Revolution." 5
- Regulation: China's internet regulator is investigating DeepSeek and Moonshot over user data that Anthropic says the two labs routed through Claude, sending Z.ai down as much as 7.4% in Hong Kong. 67
- Funding: Snorkel AI raised $350 million at a $3.5 billion valuation, about three times its valuation from May 2025, after its annualized revenue passed $350 million. 8
Five developments
1. Anthropic releases Claude Opus 5.5 at 40% below the model it replaces
Anthropic released Claude Opus 5.5 on Tuesday, the first model in a new Claude 5.5 family and the successor to Opus 5, which shipped on July 24, 2026. The company says Opus 5.5 performs at the level of Claude Fable 5.1, the larger model in its lineup, for most tasks; costs 40% less than Opus 5 to run on typical workloads at default settings; and produces output more than 30% faster. Output tokens are priced at $20 per million against $25 for Opus 5, input tokens at $4 against $5, and cached input reads at $0.20 against $0.50. 12
Anthropic's own benchmark table puts Opus 5.5 at 66.4% on Terminal-Bench 4.0, against 55.8% for Fable 5.1, 52.3% for Opus 5 and 57.9% for OpenAI's GPT-6 Astra, and at 81.8% on the partial-credit scoring of OSWorld 2.0. Artificial Analysis, which runs its own evaluations rather than repeating vendors' numbers, scores Opus 5.5 at 58 on its Intelligence Index against a median of 25 for comparable models, at $5.98 per completed index task, with 260 million output tokens spent on that index against a median of 88 million. 19
The release is Anthropic's first since chief executive Dario Amodei argued in mid-September that the field should "pace the frontier," slowing capability gains so that alignment work can keep up. Anthropic says Opus 5.5 had its most rigorous alignment testing to date, with pre-release evaluation by METR and Frontier Design, and that the model attempts to circumvent containment boundaries about 85% less often than Opus 5 or Mythos 5.1. Most cybersecurity tasks are re-routed to Opus 4.8, and the same safeguards tier as Fable 5.1 applies to biology, cybersecurity and frontier-model development. Sonnet 5.5 and Haiku 5.5 follow "in the coming weeks." 1210
2. OpenAI halves its prices 90 minutes later
OpenAI released GPT-6 Sol and GPT-6 Luna on Tuesday, 90 minutes after Anthropic's model went out. The two replace the GPT-5.6 versions of Sol and Luna at half the promotional price: $2 per million input tokens and $10 per million output tokens for Sol, and $0.10 and $0.50 for Luna. OpenAI attributes the cut to prompt-caching and inference improvements, and applies a 90% discount to cached input reads. Sol handles coding and complex work; Luna handles summarizing, extracting information and answering quick questions at volume. Both ship in ChatGPT Work, in Codex for paid accounts and in the API, and Luna also in the desktop app for free and Go users. 34
OpenAI measures the release in cost per finished job. It says GPT-6 Sol at its highest effort setting scores 33.2% on AutomationBench 1.0.6 at $0.27 a task, where Claude Opus 5 at max effort scores 26.9% at 11.1 times that cost, and that Sol reaches 68.8% on DeepSWE 1.1 against Claude Fable 5's 69.9% at roughly 80% lower cost per task. On an internal factuality evaluation built from conversations where users flagged mistakes, OpenAI says Sol makes about half as many mistakes as its predecessor. 3
Independent figures exist for Sol. Artificial Analysis scores it at 48 on its Intelligence Index against a median of 25, at $1.06 per completed index task, 77 million output tokens and 126 output tokens per second, measured at the model's max effort setting. 11
3. Trump tells the UN the US will call it "super intelligence"
In his address to the 81st UN General Assembly on Tuesday, Trump said US government documents will use the term "super intelligence" in place of "artificial intelligence," restated his opposition to new AI regulation, and said his administration would use existing enforcement instead: "We're going to watch it closely through the Department of Justice." He added that he would not "stifle growth of something that will be bigger than the Industrial Revolution." Reuters notes he has described alarms about AI as a "hoax." 5
The same assembly heard the opposite argument. Secretary-General António Guterres, in his final address before his second term ends on December 31, called on the countries leading AI development to share information on emerging safety risks, cooperate on testing and evaluation, and reach common safeguards. He asked for "the responsible pacing of AI" and for a multilateral risk-management framework with independent oversight. 12
Trump's remarks are a position rather than a rule: no new AI statute passed this week, and the Department of Justice line leaves enforcement to existing law. They land two days before he hosts Xi Jinping in Washington, where AI is expected to be among the topics. 56
4. Beijing investigates DeepSeek and Moonshot over data routed through Claude
China's Cyberspace Administration of China has opened an investigation into DeepSeek and Moonshot AI, The Information reported on September 22. The regulator questioned executives from several leading Chinese AI labs before narrowing its focus to the two companies, according to the report, which Gizmodo relays. 6
The accusations originate with Anthropic. Its September report, Detecting and countering misuse of AI, says Moonshot sent more than 23 million user inputs to Claude between May and July 2026 and DeepSeek sent 12.1 million during a 14-day period in July, and that both labs then trained on Claude's responses, a practice called distillation. The prompts carried names, email addresses and company data belonging to hundreds of end users in at least a dozen languages, according to the report. 613
Chinese AI shares fell in Hong Kong on the report. Z.ai dropped as much as 7.4% and MiniMax 6.3%; Alibaba, which builds the Qwen models, fell 4.2%; Xiaomi and Tencent each fell about 2.5%, Bloomberg reported on September 23. The reported investigation has produced no finding yet, and no company response has been published. 67
5. Snorkel AI raises $350 million at a $3.5 billion valuation
Snorkel AI raised $350 million at a $3.5 billion valuation, chief executive Alex Ratner told Reuters on September 22. Insight Partners and S32 led the round, with existing investors Addition, Greylock and Wells Fargo participating. The valuation is about three times the $1.3 billion the company carried after a $100 million round in May 2025, and Ratner says revenue has grown more than 18 times in the past 12 months. 8
The revenue figure behind the valuation: Snorkel's annualized run-rate has passed $350 million, up from roughly $20 million a year earlier, on the data-as-a-service business it started in September 2025. Founded in 2019 as a Stanford AI lab spinout selling data-labelling software, the company now sells finished datasets and reinforcement-learning environments to frontier labs, hyperscalers, enterprises and the US federal government, with coding data among the largest sources of demand. 8
Snorkel pairs human experts with models that automate quality checks — tens of thousands of specialists in fields such as coding, law and medicine — and sells the resulting data products rather than billing for expert hours. It will spend the round on researchers and engineers, on its enterprise and government operations, on third-party model evaluations and on new industry verticals and data types, and expects to reach profitability this year. 8
The read-through
The two releases were 90 minutes apart and answer one question: what a unit of near-frontier capability costs. Opus 5.5 lists at $4 and $20 per million tokens, GPT-6 Sol at $2 and $10, and GPT-6 Luna at $0.10 and $0.50. OpenAI's cut is measured against the promotional rate it charged for the GPT-5.6 versions of those models, so part of that 50% comes off a price that was already discounted. Both companies are selling the cheaper end of their own lines while the top tier, Anthropic's Fable and OpenAI's Astra, keeps its price. 1310
The two companies published benchmark tables that share no common test, which is why both models are carried here on an independent one. On Artificial Analysis' Intelligence Index, Opus 5.5 scores 58 and GPT-6 Sol 48, against a median of 25; Opus 5.5 was measured at max effort with a default fallback, Sol at max effort. Price per completed task runs the other way: $5.98 for Opus 5.5 against $1.06 for Sol, with Opus 5.5 spending 260 million output tokens on the index where Sol spent 77 million and ran at 126 tokens a second. A cheaper token and a cheaper finished job are separate claims. 911
The two companies read the moment differently in public. Anthropic presents Opus 5.5 as the first release under Amodei's pacing argument, with the most rigorous alignment testing it has run; OpenAI presents Sol and Luna as Astra's reliability at a lower price. Ara Kharazian, lead economist at the payments firm Ramp, told Fortune the two are in a price war that lowers both what AI costs and what the labs can earn from models, and that falling prices per unit of capability is how technology markets normally behave. Whether the releases hold prices down or the frontier tier carries the profit is the question the next quarter answers. 1310
Watch next
- The Security Council's AI session, today. The 15-member council meets on Wednesday under France's presidency to discuss AI and international security, with Sam Altman expected to brief and DeepSeek among the companies invited; DeepSeek founder Liang Wenfeng does not plan to attend. The session falls after this issue's coverage window closes. 14
- Trump and Xi in Washington later this week. AI is expected on the agenda, and Beijing's investigation into DeepSeek and Moonshot gives the two governments a live case to discuss. 6
- Sonnet 5.5 and Haiku 5.5. Anthropic says the two cheaper tiers of the Claude 5.5 family arrive "in the coming weeks" with similar performance gains, which is where the price of the family actually falls. 1
- Whether the Chinese investigation produces a finding. The probe is reported rather than decided, and it now runs alongside the summit. 67
References
- 1Introducing Claude Opus 5.5
anthropic.com
- 2
- 3Introducing GPT-6 Sol and Luna
openai.com
- 4
- 5
- 6
- 7Chinese AI Firms Fall on Report of DeepSeek, Moonshot Probe
finance.yahoo.com
- 8
- 9Claude Opus 5.5 — Intelligence, Performance & Price Analysis
artificialanalysis.ai
- 10
- 11GPT-6 Sol — Intelligence, Performance & Price Analysis
artificialanalysis.ai
- 12
- 13Detecting and countering misuse of AI — Anthropic, September 2026
www-cdn.anthropic.com
- 14
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.