Tech News · 2026-08-02
GPT-5.6 rewrote the code that serves it, and OpenAI cut its cheapest tier's price 80% in the same release.
GPT-5.6 Sol rewrites its own serving code, cutting Luna's price 80%1
GPT-5.6 Sol rewrote and optimized OpenAI's own production inference kernels, OpenAI says, cutting serving costs 20% and lifting token-generation efficiency more than 15% through the model's own experimentation.1 Luna, OpenAI's cheapest tier, now costs $0.20 per million input tokens and $1.20 per million output tokens, an 80% drop that undercuts Google's Gemini 3.1 Flash-Lite at equivalent price points.12 Terra fell 20% to $2 per million input tokens and $12 per million output tokens. Sol's own per-token price held, gaining only a Fast mode that runs 2.5 times faster at twice the cost.1 A lab that rewrites and prices its own serving stack turns an internal efficiency claim into a retail price the same day.
DeepSeek's V4-Flash exits preview at half its old price3
DeepSeek's own pricing page puts V4-Flash-0731's output price at $0.28 per million tokens, a cut Caixin reports as roughly 50% below the prior rate.34 The release supersedes April's preview, ships without the V4-Pro companion DeepSeek had targeted for mid-July, and still trails Moonshot's Kimi K3 on capability by Caixin's account.4 Axios frames the release as one entry in a broader price war running the same week across OpenAI, Google, and Meta.5 It quotes an unnamed OpenAI adviser: "spending tens of billions to build a slightly smarter model may buy only a temporary lead, without creating lasting pricing power."5 Two labs cutting inference prices the same week means the floor is now set by competition, not by how much headroom either company's own engineers found.
IETF bans RSA and Diffie-Hellman key exchange in TLS 1.26
RFC 10015 converts what had been non-normative guidance into a hard prohibition. TLS 1.2 and DTLS 1.2 clients and servers must not offer or select any RSA or finite-field Diffie-Hellman key-exchange cipher suite.6 The change touches more than 175 named suites and updates 17 prior IETF documents.7 The RFC cites the 2020 Raccoon timing-side-channel attack as the rationale for banning finite-field DH, rather than continuing to merely discourage it.6 Tech Times notes PCI-DSS 4.0 does not name the newly prohibited suites, so a payment system can stay card-industry compliant while running a cipher the IETF now forbids.7 Practitioners flagged the resulting bind in discussion threads. Pre-2013 embedded and government systems that cannot negotiate TLS 1.3 will need explicit configuration carve-outs.8 Some peers already hard-disconnect on any offered insecure suite, which forces the upgrade even on hardware nobody can patch.8 TLS 1.2 itself, and both TLS 1.3 branches, stay untouched. A decade of informal warnings just became something a compliance scanner can fail you on.
Gemini Robotics 2 learns new robot bodies from fewer than 200 examples9
Gemini Robotics 2 adapts to a new bi-arm robot body from typically fewer than 200 examples, DeepMind says.9 The same release ships that adaptation across five named hardware partners: Apptronik's Apollo 2 humanoid, Franka's Duo arm, Dexmate, SO101, and Trossen platforms.9 A second model in the release, Gemini Robotics ER 2, plans and coordinates task sequences that run several minutes and hundreds of decisions, evaluated on a new ASIMOV-Agentic safety benchmark DeepMind built for exactly that longer-horizon class of task.9 Engadget's independent look at the demonstrations found each one specifically trained through teleoperation, video examples, and simulation, not the general-purpose autonomy DeepMind's "real-time, fully autonomous" framing implies.10 It is the same gap that undercut Tesla's Optimus demonstrations once their teleoperators came to light.10 Fewer than 200 examples is now enough to move a trained policy onto hardware it has never seen, but nothing in the release shows the model reaching that hardware without a human demonstrating the task first.
Sources
- OpenAI · Advancing the price-performance frontier with GPT-5.6
- Simon Willison · GPT-5.6 Luna's 80% price drop
- DeepSeek API Docs · Models & Pricing
- Caixin Global · DeepSeek Releases Official V4-Flash Model as China's AI Race Intensifies
- Axios · DeepSeek's new bargain model accelerates AI's race to zero
- RFC Editor · RFC 10015: Deprecating Obsolete Key Exchange Methods in TLS 1.2 and DTLS 1.2
- Tech Times · IETF Formally Bans RSA and DHE from TLS 1.2: Over 175 Cipher Suites Off-Limits
- Hacker News · RFC 10015 discussion thread
- Google DeepMind · Gemini Robotics 2 Brings Whole Body Intelligence to Robots
- Engadget · Google's New Gemini Robotics 2 Platform Allows for "Intelligent Whole-Body Control"