Presented by
Good morning, AI enthusiasts. Anthropic just dropped a new SOTA coding model that costs 40% less to run, and OpenAI answered within hours by cutting GPT-6 Sol and Luna prices in half.
Two frontier labs racing to cut prices on the same day isn't a coincidence, so has the era of paying flagship prices for flagship results quietly ended?
In today's recap:
Anthropic ships Opus 5.5 at 40% lower cost than Opus 5
OpenAI cuts GPT-6 Sol and Luna prices in half
Wire an AI agent to GitHub without exposing credentials
DigitalOcean opens Managed Agents to the public
5 new AI tools, prompts, and more
ANTHROPIC
Anthropic drops a SOTA coding model at a lower price
Source: Anthropic
Recaply: Anthropic just introduced Claude Opus 5.5, a new state of the art model in agentic coding and computer use, that costs 40% less to run than Opus 5 while performing at the level of Claude Fable 5.1 on most tasks.
Key details:
Early testers completed a 680,000-line code migration in under a day and fixed load-time bugs across 39 of 40 pages of a test app, work that previously took engineering teams weeks with Opus 5.
Input and output tokens now cost $4 and $20 per million, 20% less than Opus 5, while cache reads used heavily in coding work drop 60% to $0.20 per million tokens.
GitHub's Chief Product Officer Mario Rodriguez said Opus 5.5 used among the fewest tokens and steps his team measured, solving more terminal tasks than Opus 5 in less than half the steps.
Opus 5.5 launched September 22, 2026 and is live now on the Claude Platform, AWS, Google Cloud, and Microsoft Azure, with Sonnet 5.5 and Haiku 5.5 following in the coming weeks.
Why it matters: Anthropic's move to cut prices while raising the performance ceiling puts pressure on every lab racing to ship a cheaper frontier model this week. With Opus 5.5 matching Fable 5.1 on much of the work builders actually do, teams no longer need flagship prices for flagship results. The timing lands the same day OpenAI cut GPT-6 Sol and Luna prices in half.
PRESENTED BY PITCH
Turn a Claude prompt into a shareable deck
Recaply AI: Most AI-made decks look AI-made: generic layouts, off-brand colors, and drafts that need a full rewrite before anyone hits send. Pitch's new MCP and API solve that problem by pulling from your real templates, so Claude (or your CRM) can build a deck that’s actually on brand and ready to send.
With Pitch, you’ll get:
A finished deck from a single Claude prompt, no manual editing required
Automatic decks triggered by CRM stage changes or form fills
A shareable link back in the same conversation
Every output matches your real template, not a generic re-theme
Connect Claude to Pitch and turn your next prompt into a deck worth sending.
OPENAI
OpenAI cuts GPT-6 Sol and Luna prices in half
Source: Artificial Analysis
Recaply: OpenAI just rolled out GPT-6 Sol and GPT-6 Luna across ChatGPT Work, Codex, and its API, cutting prices in half compared with GPT-5.6 promotional pricing while pushing both models closer to Astra-level performance.
Key details:
Sol and Luna reuse GPT-6 Astra's training methods, bringing Astra's professional work, factuality, coding, and computer use gains down to smaller, faster models built for everyday tasks.
Sol's API price dropped from $4 to $2 per million input tokens and $20 to $10 for output, while Luna fell from $0.20 to $0.10 input and $1.20 to $0.50 output.
On AutomationBench, Sol at its highest effort setting outperforms Claude Opus 5 at max effort for just 9% of Opus 5's cost per task, according to OpenAI's own benchmark data.
GPT-6 Sol and Luna are live now in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, with free and Go users getting Luna only in the desktop app.
Why it matters: OpenAI's push to make Sol and Luna cheaper while chasing Astra-level scores is a bet that most everyday work doesn't need its flagship model. Landing the same day as Anthropic's price cuts on Opus 5.5, it turns cost per task into the real competitive battleground for the next wave of agent builders. Expect more labs to follow with tiered pricing built around effort level rather than a single flat rate.
GUIDES
Wire an AI agent to GitHub without exposing credentials
Recaply: In this tutorial, you will learn how to connect a hosted coding agent to GitHub and other SaaS tools using DigitalOcean's Action Gateway, keeping every credential outside the agent's sandbox and chat transcript.
Step-by-step:
Install the doctl beta CLI, confirm "doctl agent --help" lists commands like start, attach, and auth, then run "doctl auth init" to sign in with your DigitalOcean personal access token.
Connect GitHub once for the whole team by running "doctl agent auth github" and approving access in the browser. Every session then shares that grant instead of a pasted personal token.
Write an agents.yaml spec naming your harness with "agent: claude-code", the repo to work on, and a "tools: [do.actions]" line that turns on Action Gateway for the session.
Set explicit permissions rules allowing mcp and git or gh commands, since gateway tool calls have no in-band approval prompt and fail under the default ask policy.
Validate the spec with "doctl agent validate agents.yaml", start the session, then run "doctl agent attach" to chat with the agent as it searches for and calls tools through the gateway.
Pro tip: Run "doctl agent exec -- " to check the sandbox without attaching to the full chat, useful for confirming the environment before handing the agent a real task.
DIGITALOCEAN
DigitalOcean opens Managed Agents to everyone
Source: DigitalOcean
Recaply: DigitalOcean just opened Managed Agents to the public, letting teams run OpenCode, Codex CLI, or their own harness on infrastructure that resumes paused sessions in about 300 milliseconds and reaches 16,000-plus tools.
Key details:
Action Gateway brokers credentials for every tool call at the moment it happens. OAuth tokens and API keys never reach the model or the sandbox it runs in.
The gateway connects agents to over 16,000 tools across 500-plus providers, including GitHub, Stripe, and Snowflake. In DigitalOcean's own tests, it matched agent intent to the right tool with 99.3% accuracy.
Qencode, a media processing company, built a support-triage agent on the platform. CEO Murad Mordukhay called it "a huge force-multiplier," saving the team 4 to 8 hours a week.
Managed Agents moved from private preview in August 2026 to public preview now. Billing runs per second of active CPU, so an idle agent waiting on a tool or model response costs nothing.
Why it matters: Most builders don't want to assemble session persistence, isolation, and credential handling just to run an agent past a demo. DigitalOcean is betting they'd rather pay by the second for a stack that already does it. That makes Managed Agents a real alternative to building this plumbing in house, landing the same week as several other tools chasing the same problem.
TOOLS
Trending AI Tools
🧑💻 Claude Opus 5.5 - Anthropic's new top model for agentic coding and computer use, at $4 input and $20 output per million tokens
👾 GPT-6 Sol - OpenAI's new everyday workhorse built on GPT-6 Astra's training, live in Codex and the API at $2 per million input tokens
💸 GPT-6 Luna - OpenAI's small, fast model, free in the ChatGPT desktop app and $0.10 per million input tokens in the API
🕵️ CAIRN - Cisco Talos's open-source toolkit that hunts AI-powered malware from VirusTotal metadata, with no binary downloads
📚 Firecrawl Alexandria - Knowledge library connecting agents to 100+ data providers, from research papers to real-time housing and job listings
NEWS
What Matters in AI Right Now?
Firecrawl introduced Alexandria, an AI knowledge library, after raising a $75 million Series B led by Smash Capital. The company says its internal tests showed Alexandria scoring 21% higher on answer quality than built-in web search tools.
Xiaomi MiMo open-sourced MiMo-V2.6, its largest reinforcement learning run yet, with the $2.6 million training run pushing DeepSWE performance from 58.4% to 72.6% for the Pro model.
Hacktron researchers broke into OpenAI's internal systems this summer using Anthropic's Claude, accessing private code before OpenAI patched the issue and paid a $6,500 bug bounty.
Codos launched a virtual Chief AI Officer that plans and runs a company's AI rollout, with a six-month pilot at one fintech freeing up an estimated 21% of total staff capacity.
OpenAI just announced an independent group of mathematicians to guide how it shares AI-driven math results, after its internal model resolved more than 100 open problems since late August.
SpaceXAI just shipped Grok 4.7, its most powerful model yet for coding and knowledge work, running twice as fast and at half the price of comparable models while closing in on Claude Opus 5 on agentic benchmarks.
🧡 Enjoyed this issue?
🤝 Recommend our newsletter or leave a feedback.
How'd you like today's newsletter?
Cheers, Jason
