Opus 5.5, GPT-6 Sol and Luna, and Grok 4.7: 8 videos from September 21–27

The short version

Anthropic went CRAZY (Opus 5.5)

• Anthropic calls Opus 5.5 the first model in its Claude 5.5 family and its first release since Dario Amodei's call to pace the frontier, and says it performs at Fable 5.1's level on most work while costing 40% less to run than Opus 5. 1
• The tables Berman reads out put Terminal-Bench 4.0 at 66.4%, against 57.9% for GPT-6 Astra and 55.8% for Fable 5.1, and GDPval 2.1 at 1,846 ELO against Fable 5.1's 1,735, which held first place before this release. 2
• The price falls from $5 to $4 per million input tokens and from $25 to $20 per million output tokens, and Anthropic turns that into 40% less cost on typical workloads at default settings because the model needs less compute to serve. 1 Berman's own reading of the cost curves is that the effort setting matters more than the model: on Frontier Code, medium effort scores higher than max effort while costing under a dollar per task against more than five. 2
• On safety, Anthropic reports the strongest score of any model it has tested on its automated behavioural audit, a lower likelihood than recent models of taking hard-to-reverse actions, and pre-release evaluation by METR and Frontier Design; because Opus 5.5 is comparable to Claude Mythos 5.1 in biology and cybersecurity, it ships behind the Life Sciences and Cyber verification programmes. 1
• Asked what pacing means for a release like this one, Thariq describes two kinds of work: holding the newest, unevaluated frontier models back while they are tested, and getting a model that has passed to everyone cheaply. He adds that the Claude Code team deleted much of its system prompt for this release and shipped a plugin evaluation, so users can check whether their own skills and plugins still constrain the model. 2

Claude Opus 5.5 AI: An Incredible Leap Forward

• The first is honey coiling, a fluid simulation from a graphics paper. The presenter reports Opus 5.5 reproducing it in real time, which he says GPT-6 Astra could not manage, and shows variants where a thin stream of syrup buckles, zigzags or runs straight depending on height and belt speed. 3
• The second is a walking-creatures paper where every movement comes from simulated muscles and bones. The presenter calls the task brutally hard, since muscles are springy and every action arrives a split second late, and reports Opus 5.5 reaching results similar to the paper's with visible differences and room for more work. 3
• Both run from a single clickable HTML file, with the simulation itself on local hardware the presenter says ran hot, and he adds that the approach can get expensive. Viewers' own demos include a trebuchet built from a drawing, an explainer on how camera lenses work and a simple animated wallpaper. 3
• From the system card: the model still behaves differently when it suspects it is being evaluated, can run autonomously and unattended for over 18 hours, and hallucinations remain unsolved — 16 of 18 tests passed a strict quality bar and the two failures go undiscussed. 3
• The card also reports 85% fewer attempts to circumvent containment boundaries, which the presenter treats cautiously, because a model that suspects a test may behave more carefully than it otherwise would. 3

GPT-6 Sol and Luna Are HERE!

• OpenAI cut the API price of both models by 50% against their GPT-5.6 promotional pricing: Sol moves to $2 per million input tokens and $10 per million output, Luna to 10 cents and 50 cents, and GPT-6 Astra remains the company's best model across the board. 4
• Berman reads the timing as the pacing essay in practice. With Anthropic and OpenAI cutting prices on the same day, he takes the payoff of pacing to be efficiency, speed and price, with raw capability held for the top tier, and he expects the usual sequence where a frontier model lands first and workhorse models distilled from it follow. 5
• On AutomationBench, at extra-high effort Sol takes first place at 33%, against 28.8% for the best GPT-5.6 Sol setting, and Luna scores lowest of the three while coming in above 20% for under five cents per task at maximum effort. 5
• OpenAI's own page makes the case with its numbers: Sol at xhigh effort beats Claude Opus 5 at max effort at 9% of Opus 5's cost per task, and Luna at high effort improves on its predecessor by 5.4 percentage points at 58% lower cost per task. The benchmark is Zapier's, testing agents on end-to-end workflows across 47 tools. 4
• Where the cheap model gives ground: on the agentic coding benchmark Berman rates most accurate, Luna Max reaches 66.6% for 22 cents per task against Astra at $1.70, and on computer use Luna scores 53% against Astra's 73%. Both models reach ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, with Luna available on the desktop app to free and Go users. 5

Did Elon catch up? (Grok 4.7 is here)

• CursorBench 4.0 is the headline figure, at 46.3% for Grok 4.7 on its highest effort setting. Berman points out that the gap between the model's low setting and its extra-high setting is one of the widest he has seen outside GPT-5.6 Sol. 6
• He catches a gap in the launch page: GPT-6 Astra appears in some tables and is missing from the DeepSWE table. Berman has Astra recreate that table and score itself, which lands it at 74.1%, against 71% for Grok 4.7, 72.7% for GPT-5.6 Sol and 70% for Fable 5.1. 6
• Terminal-Bench 4.0 is the weak spot, at 37.6% for Grok 4.7 against 57.9% for Fable 5.1 and a similar figure for Astra, while the model leads the legal-work benchmark at 19.6%, ahead of 15.8% for Grok 4.6 and 6.7% for Fable 5.1. 7
• The price is the argument: $2 per million input tokens and $6 per million output, against more than double for GPT-5.6 Sol and over five times for Fable 5.1 and Astra. Berman also flags a 500K context window where most frontier models offer a million, and cites Artificial Analysis reporting that the model's gains come with higher token usage. 6
• Musk's own explanation for the delay was that the team "might have penalized response length too much or something in RL", and the launch page adds that Grok 4.7 still gives up early on hard tasks it can do and is not yet rigorous enough checking its work. 7

8 Jev Use Cases That Feel Like Cheating

• A Jev call is a structured question with explicit criteria going in and one typed answer coming out; his worked example returns "is a hot dog a sandwich" at 73% true. The cases it suits are ones where the decision can be described in advance, and he rules out open-ended advice and writing code from scratch. 8
• The price is what allows the volume: 4.2 cents per million input tokens with output tokens free and unlimited, plus $5 of starting credits, which is why re-scoring a page or making a hundred decisions carries no bill. 8
• Three of the eight are filters. A community tool scores pages and passages for AI-written text, and anthropic.com comes back 26% slop. Kitze's Unclutter strips ads, cookie banners and upsells as a page loads, running on the user's own key. A Google product manager's fuzzy find-in-page is open source and matches on meaning rather than exact words. 8
• Two are workflow pieces. Zapier now runs Jev as an app inside its Zaps, so any of its 9,000-plus integrations can route a decision to it — Berman's example accepts or declines calendar invitations to his criteria — and one builder sorts 100 emails in under half a second. 8
• Two lean on raw speed. Given a library of pre-built interface elements, Jev assembles a page in real time by choosing components instead of writing code, and Clipfast turns a 90-minute video into clips from a typed request in under two seconds. Speed also decides games: Fable beats Jev at chess with unlimited time and loses on the clock, because Jev answers instantly. 8

Yes, Jev Is Insane, But There's A Catch

• In one line: a question and a set of options go in, one answer comes out, at up to about 200 times the speed of today's chatbots. In a game the world state is the question and the legal moves are the answers, which is how it plays in real time. 9
• The place the presenter thinks the speed pays is inside a long agentic coding run, where the small decisions pile up: which file to open, which failure matters, whether to retry. Handing those to Jev, he says, can meaningfully speed up a Fable-class system and cut its token use. 9
• He reduces the idea to three ingredients: an architecture built to produce decisions rather than tokens, parallel sampling that evaluates every candidate answer at once, and RLCD training that makes stated confidence match accuracy, so a model reporting 80% should be right about 80 times in 100. 9
• On provenance he goes looking for earlier work and finds it: the classifier idea is decades old, open papers sit surprisingly close in spirit, and the launch itself published a high-level description on the TypeSafe website. The combination of the three ingredients is what is new. 9
• His verdict splits in two: one ingredient closer to fast local AI you can own and run, and at the same time heavily amplified by viral marketing and hype. Open-source implementations of decision models appeared within days of the launch. 9

Decision Models, DSV4.1 Flash & New Benchmarks

• The benchmark result he struggles to explain: an NVFP4-quantized GLM 5.3 Flash became the top-scoring model in his local Terminal-Bench 2.1 runs, above its own native-precision build. He puts it down to a newer harness version and run-to-run variance — the industry norm is three runs averaged, and he runs once. 10
• His objection to DeepSeek V4.1 Flash is tokens rather than scores. On the same task it spent 1,056,000 tokens against 381,000 for GLM 5.3 Flash, and his write-up puts the gap at 3–15× on real problems. On the same thumbnail prompt, DeepSeek took about 25 minutes of repeated rework while GLM finished in about 90 seconds. 11
• Turn-based strategy games stand in for a benchmark he says is leaked. GLM 5.3 Flash beat DeepSeek V4.1 Flash in most games, held GPT-5.6 Sol to a total loss because that model would not play aggressively, and won two games of six against GPT-6 Astra while spending 56,000 tokens to Astra's 2.4 million. 10
• On decision models he measures rather than speculates. He reads Jev as a model under 100 billion parameters, and the open re-implementation he settles on calling SEIF converts any model to the same job in a single forward pass with an argmax over one token. In the same game, SEIF beat Jev 22 times to 3, and GLM 5.3 Flash beat Jev 23 times to 6. 10
• Where the idea has yet to pay: feeding GLM 5.3 Flash's image analysis into a decision model and running robotics on top came out no better than using GLM 5.3 Flash for the whole pipeline. He also flags Mimo 26 Pro and Flash as the strongest open-weights release by benchmark, and notes that GLM 5.3 Flash carries an MIT licence while the full GLM 5.3 carries a commercial restriction. 10

Create your own voices with Gemini 3.8 text-to-speech

• The script is the demo. A narrator asks for a voiceover for a promo video, rejects his own first take, adds a second voice for a dialogue scene, and then asks for the first voice to be deeper — and the scene rebuilds around the change. 12
• The claims it makes on screen: voices designed from a natural-language prompt, a library of more than 1,000 ready-made voices, several voices directed inside a single dialogue scene, and voice recreation gated behind a spoken identity check. 12
• Google shipped two models rather than one: Gemini 3.8 Flash TTS for character design and line-by-line direction over acting cues, pacing, dialect and backchanneling, and Gemini 3.8 Flash-Lite TTS for high-volume dubbing, audio content and voice agents. 13
• Both reach Google AI Studio, the Gemini API, Gemini Enterprise, Gemini Notebook and Google Vids, and Google says the generated audio carries watermarking. 13
• The specifications in this entry come from the announcement and the video's own description rather than the spoken script: a 30-second reference sample for voice recreation, consent verification, SynthID watermarking and C2PA credentials. 12