news: GPT-6, Anthropic, Agentic coding is becoming, Security keeps moving toward

Window: last 48 hours.

Quick Summary

• OpenAI shipped GPT-6 Sol and Luna, plus better GPT-6 prompt caching. The big theme is lower-cost long-running agents: shared context can get cached input-token discounts up to 90%, and Sol/Luna API prices are 50% below GPT-5.6 promo pricing. Source: https://openai.com/index/introducing-gpt-6-sol-and-luna/ and https://openai.com/index/better-prompt-caching-for-gpt-6/
• Anthropic released Claude Opus 5.5 on September 22. It says Opus 5.5 is roughly Fable 5.1-level on most work, 40% cheaper than Opus 5, and 30% faster at generation. Source: https://www.anthropic.com/claude-opus-5-5
• Agentic coding is becoming an operating system layer, not just a chat box. JetBrains announced Air as a product system for IDE agents, team workflows, and governance. Source: https://blog.jetbrains.com/blog/2026/09/22/introducing-jetbrains-air/
• Security keeps moving toward AI-versus-AI. Palo Alto Networks announced a subscription service that uses frontier and open-weight models to continuously test customer systems for vulnerabilities and attack paths. Source: https://www.axios.com/2026/09/22/palo-alto-networks-cyber-defense-ai-agents
• Check Point reported active exploitation of two high-severity vulnerabilities, including CVE-2026-85102 and CVE-2026-93616, both with CVSS 9.8. Source: https://blog.checkpoint.com/security/security-advisory-action-required-active-exploitation-of-cve-2026-85102-and-a-management-pre-authentication-vulnerability-cve-2026-93616/

AI and Models

• OpenAI introduced GPT-6 Sol and GPT-6 Luna as cheaper models in the GPT-6 family. OpenAI says GPT-6 Sol/Luna bring Astra-era gains to lower-cost tiers, with GPT-6 Sol at $2 per million input tokens and $10 per million output tokens, and GPT-6 Luna at $0.10 input and $0.50 output. Plain English: OpenAI is trying to make serious agent work cheaper enough for everyday usage, not just flagship demos. Source: https://openai.com/index/introducing-gpt-6-sol-and-luna/
• OpenAI also published better prompt caching for GPT-6. Prompt caching means the model can reuse previously processed instructions, tools, and context instead of recomputing everything each turn. OpenAI says cached input tokens can receive discounts up to 90%, and developers now get diagnostics for cache misses. Source: https://openai.com/index/better-prompt-caching-for-gpt-6/
• Anthropic released Claude Opus 5.5 on September 22. The company says it performs at Claude Fable 5.1 level on most work, costs 40% less than Opus 5, generates output over 30% faster, and has stronger prompt-injection resistance than Opus 5. Source: https://www.anthropic.com/claude-opus-5-5
• Anthropic's benchmark claims put Opus 5.5 at 66.4% on Terminal-Bench 4.0 and 54.4% on FrontierCode v1.1 main. Plain English: these are agentic coding tests where a model has to make multi-step code changes rather than answer trivia. Source: https://www.anthropic.com/claude-opus-5-5

Dev Tools and Infrastructure

• JetBrains announced Air as a system for agentic software development. It includes Air in JetBrains IDEs, Air Teams, and Air Governance. Plain English: JetBrains is packaging coding agents with review surfaces, team coordination, cost controls, and auditability. Source: https://blog.jetbrains.com/blog/2026/09/22/introducing-jetbrains-air/
• JetBrains also points to ACP, the Agent Client Protocol, as a way to connect IDEs to different agent harnesses. That matters because agent work is drifting toward interoperability: same codebase context, different model or agent runner. Source: https://blog.jetbrains.com/blog/2026/09/22/introducing-jetbrains-air/
• MiniMax's minimax-code repository describes itself as an open-source terminal coding agent powered by MiniMax, with an MIT license for first-party code. Plain English: the terminal-agent field is getting more crowded and more open-source. Source: https://github.com/MiniMax-AI/minimax-code
• Queue note: sandboxed execution environments still matter for coding agents. Sandboxes make agent work safer, reproducible, and easier to isolate from local credentials or accidental machine-wide changes. Source: local queue item x-ian-dot-so-modal-sandboxes.

Industry and Big Tech

• Dreamforce 2026 kept enterprise AI agents at the center. Salesforce coverage highlighted AIforce, agentic enterprise workflows, Slack updates, TSA AI agents, and defense AI work with Nvidia and OpenAI capabilities. Source: https://www.techradar.com/pro/live/dreamforce-2026-live-were-in-san-francisco-for-salesforces-big-ai-event
• Salesforce CEO Marc Benioff framed the shift as "not the end of software" but potentially the end of software that makes humans do all the work. Plain English: Big Tech's enterprise pitch is moving from "AI assistant" to "software that performs work across systems." Source: https://www.techradar.com/pro/live/dreamforce-2026-live-were-in-san-francisco-for-salesforces-big-ai-event
• VC signal: Axios Pro Rata reported Cyera raised a $400 million Series G extension at a nearly $12 billion post-money valuation, and Adlib Software agreed to acquire Paperbox, an agentic AI insurance platform. Plain English: security and vertical AI workflows are still getting funded even as generic AI wrappers feel crowded. Source: https://www.axios.com/newsletters/axios-pro-rata-8b88c78a-23fd-4eb9-91bf-466e46447815
• Queue note: Meta-style AI-native pods and AI adoption targets are worth watching. The real signal is leverage, flattening, and measurement; the risk is bad incentives if teams optimize for visible AI usage instead of durable quality. Source: local queue item meta-ai-native-pods-ai-adoption-mandates.

Security

• Palo Alto Networks announced a service that uses a mix of frontier models, including Anthropic Mythos 5 and OpenAI GPT-5.6-Cyber, plus open-weight models, to continuously test web apps, APIs, cloud infrastructure, source repos, and network assets. Source: https://www.axios.com/2026/09/22/palo-alto-networks-cyber-defense-ai-agents
• Palo Alto told Axios that in its own tests no single AI model caught more than 40% of vulnerabilities in a complex environment. Plain English: multi-model security stacks may become normal because each model misses different things. Source: https://www.axios.com/2026/09/22/palo-alto-networks-cyber-defense-ai-agents
• Check Point reported active exploitation of CVE-2026-85102, a pre-authentication remote code execution flaw in VPN certificate handling, and CVE-2026-93616, a pre-authentication path traversal flaw in Security Management. Both are listed as CVSS 9.8. Source: https://blog.checkpoint.com/security/security-advisory-action-required-active-exploitation-of-cve-2026-85102-and-a-management-pre-authentication-vulnerability-cve-2026-93616/
• Check Point says fixes are available and affected customers should install them immediately. Plain English: this is patch-now territory, especially for internet-facing gateway and management surfaces. Source: https://blog.checkpoint.com/security/security-advisory-action-required-active-exploitation-of-cve-2026-85102-and-a-management-pre-authentication-vulnerability-cve-2026-93616/

Creator Watchlist

• Theo / t3.gg had the freshest relevant watchlist activity. A recent clip, "Jev is incredible. I'm scared people are using it wrong," argues that Jev should not be treated like a chatbot replacement; the comments frame it more like a deterministic decision layer or "smart if statement." Source: https://www.24vids.com/video/jev-is-incredible-im-scared-people-are-using-it-wrong-p-tw-2101857305570721847
• Theo also surfaced a "Grok 4.7 is a disappointing release" clip within the window. Signal: creator-dev attention is still sharply focused on whether new models actually improve real coding and visual/task performance, not just benchmark charts. Source: https://www.24vids.com/video/grok-47-is-a-disappointing-release-p-tw-2102493204641939663
• The other watchlist sources did not surface a strong last-48-hour transcript item from available local checks. I skipped them rather than forcing stale content.

Takeaways

• The agent cost curve is the real story. OpenAI's cache tooling and Anthropic's Opus 5.5 pricing both point to the same bottleneck: long-running agents are only useful at scale if repeated context is cheap. Sources: https://openai.com/index/better-prompt-caching-for-gpt-6/ and https://www.anthropic.com/claude-opus-5-5
• Agentic coding is moving from individual tools to managed systems. JetBrains Air, MiniMax Code, and the sandboxing queue note all point toward the same stack: agent runner, permissions, context, review, governance, and reproducibility. Sources: https://blog.jetbrains.com/blog/2026/09/22/introducing-jetbrains-air/ and https://github.com/MiniMax-AI/minimax-code
• Security is shifting from "scan periodically" to "continuous adversarial testing." Palo Alto's AI-vs-AI service and Check Point's active exploit advisory make the same point from opposite sides: attackers and defenders are both speeding up. Sources: https://www.axios.com/2026/09/22/palo-alto-networks-cyber-defense-ai-agents and https://blog.checkpoint.com/security/security-advisory-action-required-active-exploitation-of-cve-2026-85102-and-a-management-pre-authentication-vulnerability-cve-2026-93616/
• For Deval: keep biasing toward fundamentals plus tool leverage. Learn the systems layer underneath agents: caching, sandboxing, evals, workflow design, and security boundaries. That is where the durable engineering edge is forming.