One-Line Summary

OpenAI cuts mid-tier model costs in half and Anthropic counters with Opus 5.5, while Snorkel AI raises $350M—shifting AI economics toward cheaper tasks and better training data.

Big Tech

OpenAI cuts GPT-6 Sol and Luna API prices by 50%

OpenAI releases two GPT-6 models—Sol and Luna—designed to bring frontier-level capability to everyday work at much lower cost, with Sol priced at $2 per million input tokens and $10 per million output tokens, and Luna at $0.10/$0.50; both are about 50% cheaper than their GPT-5.6 counterparts and are available in ChatGPT Work and Codex for paid tiers, with Luna also in the desktop app for Free and Go users and both models exposed in the API as gpt-6-sol and gpt-6-luna. ¹

OpenAI attributes the cuts to improvements in caching and inference, adds a 90% discount on cached input-token reads, and ships a Prompt Caching Dashboard and diagnostics to help teams raise cache hit rates; the company also reports internal gains in factual reliability, with GPT-6 Sol making roughly half as many mistakes as GPT-5.6 Sol in a de-identified ChatGPT evaluation. ¹

On OpenAI’s task-focused evaluations, Sol emphasizes lower cost per completed task: on AutomationBench 1.0.6, Sol (xhigh) scores 33.2% at $0.27 per task versus Opus 5 (max) at 26.9% costing 11.1× more; Claude Fable 5.1 with Opus 5 fallback posts 31.4% at more than 8.9× Sol’s cost, while Sol (max) hits 56.4% on Agents’ Last Exam at 60% lower cost, and 68.8% on DeepSWE 1.1—within 1.1 points of Fable 5 (xhigh)—at about 80% lower cost. ¹

VentureBeat notes that Anthropic’s Opus 5.5, announced the same morning, changes the comparison set; Opus 5.5 lists at $4 per million input and $20 per million output tokens, and there is not yet a same-harness public result establishing whether Sol or Opus 5.5 delivers a lower cost per successful task. ²

Anthropic launches Claude Opus 5.5 at lower price

Anthropic, the company behind the Claude assistant, releases Claude Opus 5.5, which it prices at $4 per million input tokens and $20 per million output—20% below Opus 5—and says it performs on par with Fable 5.1 on most tasks while costing about 40% less to run on typical workloads; Anthropic also highlights speed, clearer writing, stronger internal safety scores, and an 85% lower likelihood of attempting to bypass safeguards versus Opus 5 or Mythos 5.1. ³

Opus 5.5 is available across Anthropic’s platforms, including AWS, Google Cloud, and Microsoft Azure, with API access as claude-opus-5-5 and $0.20 per million tokens for cache reads; Anthropic says it involved outside evaluators ahead of release and is raising five-hour usage limits for Pro, Max, and Team subscribers. ⁴

Industry & Biz

Snorkel AI raises $350M at a $3.5B valuation as data needs grow

Snorkel AI, which helps enterprises build complex training data and simulated environments for AI systems, raises $350 million at a $3.5 billion valuation; CEO Alex Ratner tells Reuters the company’s annualized revenue run-rate is about $350 million and that Snorkel expects to reach profitability this year, with the round led by Insight Partners and S32. ⁵

Reuters reports that demand from frontier labs for more complex data is accelerating, underscoring how budgets are shifting from only compute to also data quality and scale—positioning data tooling as a direct lever on model performance. ⁵

New Tools

OpenAI prompt caching upgrades for GPT-6

OpenAI improves prompt caching for GPT-6 with a 90% discount on cached input reads, a Prompt Caching Dashboard, diagnostics to surface missed cache opportunities, the ability to adjust reasoning effort and tool availability without breaking cache, and explicit breakpoints to choose where cacheable prefixes end. ¹

OpenAI says GitHub observed that these caching changes cut the share of tokens requiring fresh processing by more than 50% across billions of Copilot requests, helping speed responses and lower effective costs on long-running agent workflows. ¹

Community Pulse

Hacker News (931↑) — Excitement about sharply lower costs is tempered by concerns over censorship, occasional regressions, and provider-by-provider token accounting. ⁶

"Wow I cannot believe Luna is getting even cheaper. IMO this is the model that is going to change the world. Everyone said tokens were too expensive but these are getting close to free while still having fantastic performance." — Hacker News ⁶

"Except for the censorship. We use it for massive data crunching, and roughly 5-8% (depending on the day) gets censored and doesn't get a response. We switched to Mimo 2.6, which is relatively better. For censored stuff, we use Sonnet and OpenAI Nano models." — Hacker News ⁶

Hacker News (982↑) — Debate centers on whether efficiency gains represent frontier-pushing progress versus practical slowdowns from over-sensitive safety classifiers that block legitimate work. ⁷

"Well yes, If you magically found a way to reduce a models energy usage by 50%, the thing that would happen immediately after is a doubling of a training scale and of the test time compute assigned for a given budget. I don’t see how that wouldn’t be considered pushing the frontier. Given that scaling is the one thing that has been bringing us closer and closer to AGI, a sudden 2x increase in scaling laws would definitely not be considered pacing. You seem to see a contradiction where I don’t." — Hacker News ⁷

"[cyber] classifier is incredibly sensitive with Opus 5.5 I cannot complete any embedded/driver/system-level tasks. Quite literally not a single task was able to complete today without getting flagged for [cyber], and what's more annoying is their narrow definition of what a cybersecurity specialist should be preventing me from getting an exception.." — Hacker News ⁷

What This Means for You

• Budgets can stretch further: Sol at $2/$10 and Luna at $0.10/$0.50 significantly lowers routine task costs, and GPT-6’s 90% cached-read discount makes long prompts and repeated instructions cheaper—useful for document-heavy workflows and agents that replay context. ¹

• Optimize for cost per completed task, not just token price: VentureBeat underscores that cross-vendor head-to-heads are still unsettled post–Opus 5.5; your real savings depend on cache hit rates, effort settings, and whether a cheaper model can finish the job without retries. Pilot with a fixed task set and measure outcomes. ²

• Expect stricter safety boundaries on some providers: Anthropic is emphasizing safeguards; plan fallbacks for low-level or security-adjacent tasks and validate whether filters block legitimate use cases in your domain before committing. ⁴

• Data quality is a new lever: Snorkel’s $350M raise and $350M run-rate signal that training and simulation data are becoming strategic spend; non-technical teams can contribute by defining labeling guidelines, edge cases, and “done” criteria that improve dataset quality. ⁵

Action Items

1. Try GPT-6 Luna for bulk summaries: In the ChatGPT desktop app, switch to GPT-6 Luna for document summarization or extraction and compare speed/clarity against your current model on 2–3 real tasks.
2. Run a Sol vs. current-model check on one workflow: If you use ChatGPT Work or Codex, run the same coding or analysis prompt with GPT-6 Sol and your current default, then decide which model becomes your new team default for that task.
3. Test Claude Opus 5.5 on a sensitive task: In Claude, try one real workflow that previously hit safety limits and see if Opus 5.5’s behavior, tone, and results meet your needs before scaling up.
4. Draft a simple routing policy: Write a one-page guide for your team: Luna for high-volume routine steps, Sol for complex work that repeats, Astra only when top-tier reasoning is required; include when to use fallbacks if a provider blocks a task.