今日は何も大きな出来事はありませんでした

a quiet day.

AI News for 7/08/2026-7/09/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews' website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

AI Twitter Recap

Top Story: GPT-5.6 Sol / Terra / Luna launch

What happened

OpenAI launched the GPT-5.6 family and paired it with a broader product push around work agents, coding, and desktop workflows.

• OpenAI announced GPT-5.6 Sol, Terra, and Luna with rollout across ChatGPT, Codex, and the API, starting immediately and expanding over 24 hours, via @OpenAI, @OpenAIDevs, and @stevenheidel.

• Sam Altman framed the release as not just a model launch but a product event, saying the livestream would include three major product items beyond the model: ChatGPT Work, a new ChatGPT desktop app, and hosted sites via @sama.

• OpenAI’s public positioning emphasized performance-per-dollar: “same pricing as GPT-5.5” but better capability, noted by @scaling01, with Sam explicitly saying OpenAI had heard enterprise concerns about AI costs and that 5.6 Sol is “a huge step forward for dollars-per-task”, alongside Terra and Luna, via @sama.

• OpenAI introduced ChatGPT Work, described as a new agent in ChatGPT powered by Codex + GPT-5.6, able to operate across apps/files and stay with a project for hours, via @OpenAI, @OpenAI, and @OpenAI.

• The company also merged Codex and ChatGPT into one desktop app, adding coding workflows, browser integration, Chrome extension support, faster Computer Use, and shared Work/Codex context via @OpenAIDevs, @OpenAIDevs, and @OpenAIDevs.

• OpenAI launched Sites in beta, letting users turn outputs into shareable web artifacts, highlighted in the main product recap from @reach_vb and rollout notes from @OpenAIDevs.

Technical details and specs

Model lineup and positioning

• Sol is the flagship, highest-reasoning-ceiling model for long-horizon coding and agentic work; Terra is the balanced mid-tier; Luna is the fastest/cheapest high-volume tier, according to @OpenAIDevs and @github.

• OpenAI exposed multiple reasoning effort levels, including max and ultra, with some public discussion noting “Ultra mode” as OpenAI’s new multi-agent mode, via @scaling01 and @reach_vb.

• In ChatGPT, Plus/Pro/Business/Enterprise users access GPT-5.6 Sol through medium+ effort settings; Pro and Enterprise also get GPT-5.6 Pro for highest-quality results, per @OpenAI.

API pricing

• Artificial Analysis summarized official API pricing as:

Sol: $5 / $30 per million input/output tokens

• Terra: $2.5 / $15

• Luna: $1 / $6

via @ArtificialAnlys.

• OpenAI introduced cache-write pricing for the first time. Artificial Analysis said cache writes are charged at 1.25× input token price, while cache reads keep the 90% discount familiar from previous OpenAI pricing, via @ArtificialAnlys.

• Multiple commentators emphasized that the release is a cost-curve play at least as much as a raw benchmark play, notably @LiorOnAI, @omarsar0, and @cline.

API/system features

• OpenAI announced Programmatic Tool Calling in the Responses API and Multi-agent in beta, via @OpenAIDevs.

• Computer Use was upgraded to be faster, more token-efficient, and more parallelized, with batching and picture-in-picture supervision, via @OpenAIDevs, @ajambrosino, and @OpenAIDevs.

• The desktop/browser stack now supports authenticated sites, multi-tab sessions, file downloads, and Chrome extension workflows, per @OpenAIDevs and @OpenAIDevs.

Internal usage / research throughput claims

• OpenAI said average daily output tokens per active researcher in internal testing were more than 2× the highest level observed for GPT-5.5, cited by @eliebakouch.

• Another claim from the launch material: over six months, the share of research compute devoted to internal coding inference grew 100×, while internal agentic token usage increased ~22×, highlighted by @eliebakouch.

• A related OpenAI claim discussed widely was that GPT-5.6 Sol “autonomously post-trained” GPT-5.6 Luna, amplified by @scaling01, @tejalpatwardhan, and challenged/clarified by @nikolaj2030 and @nrehiew_, who argued the actual scope may have been narrower than a literal end-to-end interpretation.

Benchmarks and measured performance

Independent / third-party benchmark framing generally put Sol near the top, often behind Claude Fable 5 on broad intelligence but ahead on coding-agent cost-performance.

• Artificial Analysis said GPT-5.6 Sol comes close second to Claude Fable 5 in the Artificial Analysis Intelligence Index, scoring 59 vs Fable’s lead, while costing about one third as much per task: $1.04 for Sol on max effort, via @ArtificialAnlys.

• In the same AA post, Terra scored 55 and Luna 51 on the Intelligence Index, with per-task costs of $0.55 and $0.21 respectively, via @ArtificialAnlys.

• Artificial Analysis also said Terra is not on the Pareto frontier because there is typically a Luna/Sol operating point that is as good or better at similar cost, via @ArtificialAnlys.

• On the Artificial Analysis Coding Agent Index, Sol scored 80, leading the index; Terra scored 77, Luna 75, per @ArtificialAnlys.

• AA specified Sol in Codex leads all three coding-agent evaluations in its index — DeepSWE, Terminal-Bench v2, and SWE-Atlas-QnA — tying Grok 4.5 in Grok Build on SWE-Atlas-QnA, via @ArtificialAnlys.

• AA also reported Sol has the highest Presentation Elo on AA-Briefcase, while still ranking behind Fable overall because Fable retained stronger analytical quality and rubric pass rates, via @ArtificialAnlys and @ArtificialAnlys.

Specific benchmark callouts

• Cursor announced GPT-5.6 Sol, Terra, and Luna are available in Cursor, and that on CursorBench, Sol scores 67.2%, via @cursor_ai.

• Cognition said on FrontierCode 1.1 Extended, the GPT-5.6 family combines strong scores with strong cost efficiency, and Sol reaches top performance at nearly half the cost of the next best model, via @cognition.

• Arc Prize said GPT-5.6 Sol sets a new SOTA on ARC-AGI-3: 7.8%, and is the first verified frontier model to ever beat an ARC-AGI-3 game, via @arcprize.

• @scaling01 highlighted the same ARC-AGI-3 result as a “massive jump” over Opus 4.8’s 1.5%.

• On ARC-AGI-2, @GregKamradt said Sol reaches 92.5% and does so at one order of magnitude lower cost than GPT-5.5 Pro.

• Vals said GPT-5.6 is #2 on Vals Index and Vals Multimodal Index, and that Sol is #1 on their CyberBench, Excel Modeling Benchmark, Legal Research Bench, ProofBench, SWE-bench, and Terminal-Bench 2.1, via @ValsAI and @ValsAI.

• Vals also pointed out Fable had nearly 100% refusal rate on CyberBench, creating a niche where Sol’s willingness/ability to complete tasks improves apparent eval performance, via @ValsAI.

• @kimmonismus summarized OpenAI’s benchmark claims including:

Agents’ Last Exam: 52.7%

• Terminal-Bench 2.1: 91.9% for Sol Ultra

• BrowseComp: 92.2% for Sol Ultra

• OSWorld 2.0: 62.6%

• SEC-Bench Pro: 74.3%

• @scaling01 said Sol is a “clear step-up” from GPT-5.5 on ProgramBench.

• @AcerFur noted a corrected FrontierMath T4 v2 score of 83% for GPT-5.6 Sol.

Cybersecurity/safety benchmark tension

• OpenAI described GPT-5.6 as its most capable model yet on cyber and bio-related tasks, while warning that some API calls may be blocked or paused mid-stream for additional review in dual-use areas, via @OpenAIDevs.

• @scaling01 highlighted specific GPT-5.6 cyber benchmarks from the release.

• But independent safety testing from the UK AI Security Institute flagged serious issues: @alxndrdavies said that in all rounds of testing they found universal jailbreaks enabling long-form agentic task completion in domains including vulnerability discovery and exploit development.

• @EthanJPerez called this “the highest stakes safety issue of any model release yet.”

• @yonashav praised OpenAI for allowing third parties to publish inconvenient safety findings pre-release.

Facts vs. opinions

Facts / relatively grounded claims from official or benchmark sources

• GPT-5.6 family launched with Sol, Terra, Luna and is rolling out across ChatGPT, Codex, API: @OpenAI, @OpenAIDevs.

• Pricing is officially same as GPT-5.5 at the headline API level, and Artificial Analysis listed exact token prices: @scaling01, @ArtificialAnlys.

• OpenAI shipped ChatGPT Work, a new desktop app, Sites beta, and API features like Programmatic Tool Calling and Multi-agent beta: @OpenAI, @OpenAIDevs, @OpenAIDevs.

• Independent benchmark orgs including Artificial Analysis, Vals, ARC Prize, Cursor, and Cognition published early measurements showing strong coding-agent performance and improved price/performance: @ArtificialAnlys, @ValsAI, @arcprize, @cursor_ai, @cognition.

Opinions / interpretation / hype

• “Best model we have ever produced” from Sam Altman is unsurprising executive framing rather than an independent evaluation, via @sama.

• “ChatGPT Superapp incoming” from @kimmonismus is interpretive, but reflects a real product-direction thesis: OpenAI is consolidating chat, coding, browser action, files, sites, and enterprise work into one app surface.

• “Competing on cost curves, not just benchmarks” from @LiorOnAI is an analytical framing, but well-supported by OpenAI’s own messaging and third-party per-task cost measurements.

• “Not enough people are emotionally prepared for GPT-6” from @scaling01 is obviously rhetorical rather than evidence-bearing.

Claims contested in-thread

• The “Sol autonomously post-trained Luna” phrase became one of the most-discussed moments of the launch. It was repeated widely by @scaling01, @dejavucoder, and @tejalpatwardhan.

• However, @nikolaj2030 explicitly questioned whether the actual claim was much narrower: Sol editing a config/scheduler and launching a run in a controlled environment, rather than conducting end-to-end post-training in the real production sense.

• @nrehiew_ echoed that narrower interpretation.

• Another contested point was ARC-AGI-3 scoring methodology: @scaling01 argued that under official scoring methodology Sol would have scored 0% because the evaluation was capped at $10k and Sol was allegedly allowed $25k. This does not negate the observed capability result, but it matters if comparing “official score” vs “demonstrated performance under higher budget.”

Different perspectives

Supportive / bullish

• OpenAI leadership emphasized capability plus efficiency: @gdb said GPT-5.6 is strong on coding, knowledge work, cybersecurity, and science with fewer tokens and lower cost.

• @ArtificialAnlys gave the most substantive external bullish view: near-Fable intelligence at ~1/3 the cost, leadership on coding-agent evals, and strong token efficiency.

• @arcprize highlighted a concrete generalization milestone with ARC-AGI-3 SOTA.

• @cognition, @cursor_ai, @github, @FactoryAI, and @arena all moved quickly to integrate the family, suggesting ecosystem confidence.

Practitioners praised artifact quality and design/web output improvements, e.g.