GPT-6 Sol and Luna half the price and on Bedrock. How to cut AI costs now.

Key AI news

GPT-6 Sol and Luna: price halved and launch on Bedrock

OpenAI priced GPT-6 Sol and GPT-6 Luna much lower. According to Simon Willison, Luna costs 0.10 USD per million input tokens and 0.50 USD per million output tokens. Sol is 2 USD and 10 USD. That is about half of the GPT-5.6 counterparts at promo price. The models are now generally available on Amazon Bedrock. Bedrock supports explicit prompt caching, IAM controls, and VPC access. OpenAI reports fewer factual errors for Sol than for 5.6 Sol in its own tests. Those are vendor claims. For you this means: You route volume jobs to Luna and keep complex cases for Sol, with cache on reusable instructions.¹².

Anthropic cuts Opus 5.5. Test “max” before rollout.

Anthropic reduces Opus 5.5 to 4 USD per million input tokens and 20 USD per million output tokens. Cache reads drop by 60 percent. That is a 20 percent cut versus Opus 5.0 to 5.1. Simon Willison reports that Opus 5.5 at “max” thinking level twice produced no answer on his SVG test because it exhausted the maximum output length. The two runs each took almost 20 minutes and cost 2.56 USD each. Single tester result, not a broad evaluation. If you plan Opus 5.5 “max”, simulate your prompt realistically and measure abort rates. Sonnet 5.5 and Haiku 5.5 are “coming soon” from Anthropic.¹.

Meta and Singapore Police act against scam networks

Meta reports extensive takedowns with the Singapore Police Force. January to June 2026: over 113,000 fraudulent entities and pages on Facebook and Instagram. June: 33,600 more in an operation against e-commerce fraud. July: proactively disabled over 3.6 million “shell pages” before activation. For you this means: You enable 2FA and passkeys for company accounts, audit ad accounts for impersonated pages, and document internal approvals. Figures come from the vendor.³.

Highlights for your workday

• You cut runtime costs right away when you split tasks: classification, extraction, and short answers to GPT-6 Luna. Longer analysis, code reviews, and tool chains to GPT-6 Sol. Mark policies, tools, and schemas as prompt cache on Bedrock. You pay less for recurring context this way. Performance claims for Sol are vendor claims.²¹.

• Swap models without downtime. Together AI describes canary rollouts with staged traffic ramps, metric gates, and automatic rollback. You define measurable gates like error rate, latency, and escalations, run 1 to 5 percent, check, then increase.⁴.

• Confidential data stays protected, even in production. NVIDIA shows confidential computing for performant LLM inference in memory-encrypted environments with confidential GPUs. If you process health, finance, or customer data, plan this mode for new AI workloads..

Tools and updates

• RetroChimera: open Microsoft framework and weights for retrosynthesis. Combines two models and learns a re-ranking. Useful for early route evaluation in pharma and materials. Chemistry data and expert review remain prerequisites.⁵.
• Industrial safety in seconds: Tata Elxsi describes IRIS on AWS. Frames are prefiltered at the edge, only relevant images and metadata flow into a serverless pipeline with real-time alerts. For operators of many cameras with lean control rooms.⁶.
• Rebalancer by Meta: open-source library for assignment and placement problems. Separation of specification and solution, heuristics or MIP. Usable for resource, task, or traffic distribution.⁷.
• Regional AI in SEA: NVIDIA Nemotron open models are used by agencies and companies. Example: ThaiLLM adapts Nemotron 3 Nano for legal applications, powers the legal assistant Thanoy with around 43,000 users. For local language applications.⁸.

Try in five minutes

A small security gain: introduce dependency cooldowns.

1) Decide that production dependencies are only updated x days after release, about 72 hours. Document the rule in the PR template.⁹. 2) Extend your CI so that automatic upgrades to production are allowed only after the cooldown expires. Staging can test earlier.⁹. 3) Pick a top dependency and ship the next version bump to staging, not production. Note release date and earliest production date in the ticket.⁹.

Expected result: less risk from compromised releases. Check criterion: no package goes to production without an expired cooldown.⁹.

A good find

John Platt on ERA from Google: many science problems can be framed as “scorable” tasks. ERA runs evolutionary steps over a search tree and uses an optimization rule to pick promising branches. Two practice points: first, design a reliable score function. Then check results for Goodhart effects and start conservatively with simple baselines like linear regression or SVM. Transferable to internal optimization tasks with a clear objective. Based on the show notes.¹⁰.

Sources

1. Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war (simonwillison.net)
2. Bring more intelligence to everyday work with GPT-6 Sol and GPT-6 Luna on Amazon Bedrock (aws.amazon.com)
3. Meta Takes Action on 3.7 Million Accounts, Pages and Content In Partnership With Singapore Police Force (about.fb.com)
4. Canary rollouts: upgrade models in production without downtime (together.ai)
5. Improving synthesis prediction of small molecules at scale with RetroChimera (microsoft.com)
6. How Tata Elxsi detects industrial safety risks in seconds on AWS (aws.amazon.com)
7. Open-Sourcing Rebalancer: A Generic, High-Performance Library for Solving Assignment Problems (engineering.fb.com)
8. At AI Day Singapore, NVIDIA and Partners Showcase AI Advancements Across Southeast Asia (blogs.nvidia.com)
9. Be alert: targeted attacks on prominent Rustaceans (simonwillison.net)
10. 🔬 An Oscar, Two Asteroids, and the Algorithm in Your sklearn: John Platt on AI for Science (latent.space)