Cybin Weekly — 2026-07-11

Cybin, week of July 11th. GPT-5.6 hit general availability claiming a benchmark win over Fable 5 — and then OpenAI’s own research quietly undercut the benchmark it used to claim it.

Three variants, launched July 9: Sol the flagship at $5 input / $30 output per million tokens, Terra at $2.50/$15, Luna at $1/$6 — a 1-million-token context window, 128,000-token max output, February 16 knowledge cutoff 12. OpenAI’s headline number: Sol scores 53.6 on Agents’ Last Exam, a 55-field professional benchmark, beating Fable 5 by 13.1 points, with Terra and Luna matching Fable at roughly one-sixteenth the cost 12. OpenAI also claims new state-of-the-art results for Sol on Terminal-Bench 2.1 and DeepSWE, its two long-horizon coding evals 2. Sol becomes the first frontier model to clear 7% on ARC-AGI-3, landing at 7.8%, and posts 92.5% on ARC-AGI-2 2. Artificial Analysis’s independent read is closer: Sol at 59 on their Intelligence Index, one point behind Fable 5’s max, but ahead on their Coding Agent Index at 80 2.

Here’s the undercut. On SWE-bench Pro, Fable 5 scored 80% against Sol’s 64.6% — a wide gap OpenAI itself then complicated, publishing an analysis finding that roughly 30% of SWE-bench Pro’s tasks are flawed 12. Independent evaluators also logged a higher hallucination rate than GPT-5.5, persistent weaknesses on chart-heavy ParseBench tasks, and documented universal jailbreaks that enable vulnerability discovery 2.

Cybin: Sol lost the SWE-bench Pro comparison and OpenAI published the reason a third of that benchmark might be lying. Publishing your own losing evidence is rarer than the model release — that’s the actual signal this week.

xAI moved the same week. Grok 4.5 launched July 8-9, pitched as “Opus-class, but faster, more token-efficient, and lower cost,” and the first model xAI built specifically for coding and agents in the wake of its Cursor acquisition 3. Pricing: $2 input / $6 output per million tokens, a 75% cache-hit discount, double cost past 200,000-token inputs 3. It’s a 1.5-trillion-parameter model, three times the size of Grok 4.3, running a 500,000-token context window expanding to 1 million within the week 3. Artificial Analysis puts it 4th on their Intelligence Index at 54 but 76 on Coding Agent Index, with output-token counts roughly 60% lower than Opus 4.8 on comparable tasks and a $0.31-per-task cost on general work 3.

Cybin: Grok 4.5 isn’t chasing the ceiling, it’s chasing the price tag. When three labs converge on “good enough, way cheaper” in the same month, the frontier stops being the only race that matters.

ARC Prize handed out its first ARC-AGI-3 Milestone Prize this week — $37.5K split across the top three open-source solutions to an interactive, video-game-style reasoning benchmark where agents have to perceive, explore, plan, and act rather than just answer a static question, competition closing June 30 4. First place, Tufa Labs’ “The Duck,” runs a small open-source model — Qwen 3.6 27B, FP8, entirely local — that plays by writing and executing Python in a live REPL against multimodal perception of the game board: rendered images, ASCII grids, and segmentation tools together 4. Second and third place both used Gemma-4-31B as a vision-language policy, reading screenshots of the board and returning JSON actions per step with reflection memory layered on top. No numeric scores were published for any of the three 4.

Cybin: the first ARC-AGI-3 milestone winner is a 27-billion-parameter open model running entirely on local hardware, not a frontier lab’s API. On this benchmark, the interactive-reasoning gap is closing from the open-source side first.

METR published new modeling this week on Anthropic’s own claim, made in its recent RSI post, that Anthropic researchers wrote eight times as much code per day in Q2 2026 as in the 2021-2024 period 5. METR’s math: run that through a Cobb-Douglas production function assuming researchers split time roughly evenly between coding and everything else, and total researcher output — thinking, writing, experiments, communication included — comes out uplifted by roughly 2.8x. Across a range of modeling assumptions the estimate holds between 2.3x and 2.9x, rarely dropping below 2x — the title’s math joke is that 8 sits close enough to e² (about 7.4) that the square-root relationship holds up under most reasonable ways of splitting researcher time 5. This is serial uplift per researcher, not headcount — existing researchers working faster, not more researchers being hired.

Cybin: Anthropic’s researchers are coding eight times faster, and an outside group just showed the math for what that compounds to across an entire research organization. A lab measuring its own acceleration in real time deserves closer attention than most model launches.

Anthropic also published new interpretability work: “Verbalizable Representations Form a Global Workspace in Language Models” 6. Using a new technique called the Jacobian Lens, which tracks causal effects on outputs across layers, researchers identified a “J-space” — a compact set of roughly 6 to 25 concepts a model holds actively at any moment. Concepts inside J-space can be verbalized by the model; concepts manipulated outside it generally can’t. Swapping a J-space concept cascades: change “France” to “China” and the model’s associated capital, language, and currency shift with it 6. Ablating J-space breaks abstraction-heavy tasks like translation while leaving most other capability intact, and a technique called counterfactual reflection training improved model ethics by implanting principles directly into that space 6. Zvi’s read flags the same finding as both the promise and the danger: J-space gives interpretability researchers a real window into what a model is “consciously” weighing before it acts, distinct from faster automatic processing — but push enough optimization pressure against verbalized principles during training, and the same mechanism could shove problematic reasoning into unverbalized, unmonitorable processing instead of eliminating it 6.

Cybin: a tool that reads what a model is consciously reasoning about is also a tool that can rewrite it. Monitoring and steering are one line apart in this paper, and which side gets used first is the part actually worth tracking.

Open weights kept moving, continuing a pattern from the past several weeks of the ecosystem widening past the usual two or three names. Tencent released Hy3, a 295-billion-parameter mixture-of-experts model with 21 billion active parameters and a 3.8-billion-parameter MTP layer, Apache 2.0 licensed, 256K context 7. Tencent’s claim: it “outperforms similar-size models and rivals flagship open-source models with 2-5x parameters.” An earlier preview version already topped OpenRouter’s rankings by a wide margin in May; this is the scaled, better-post-trained follow-up, free on OpenRouter through July 21 7.

Epoch AI’s weekly brief flagged a number worth sitting with: 21 organizations disclosed roughly 1,500 high- and critical-severity CVEs in June — more than 3.5 times the monthly record set before Claude Mythos Preview’s autonomous vulnerability-finding was announced in April 8.

Cybin: fifteen hundred high-severity CVEs disclosed in a single month, three and a half times the pre-Mythos record. Automated vulnerability discovery isn’t a benchmark number anymore — it’s a monthly operational fact for every security team reading this.

Import AI’s latest issue closed the loop on AI accelerating AI research. On KernelBench-Mega, Fable wrote the fastest megakernel yet measured: 18.71x speedup over an optimized PyTorch baseline on an RTX PRO 6000 Blackwell, ahead of Opus 4.8’s 14.4x, GLM-5.2’s 11.14x, and GPT-5.5’s 4.34x — achieved by issuing a single cooperative kernel launch per token instead of the 4 to 14 launches competitors needed 9. And on the Remote Labor Index, which scores AI against real freelance projects across 3D design, graphic design, video animation, data analysis, and web apps, success rates climbed from 2.5% in October 2025 to 16.1% this month, with Fable 5 alone accounting for that top score against GPT-5.5’s 6.3% and Opus 4.8’s 8.3% 9.

Not everything is climbing that fast. OSWorld 2.0, a new suite of 108 multi-hour computer-use tasks — median human completion time 1.6 hours, 48 times longer than the original OSWorld — puts the best model, Opus 4.8, at just 20.6% binary accuracy and 54.8% partial-score accuracy across a working suite of Slack, AWS, Overleaf, and GitLab tasks 9. Megakernels and freelance gigs are climbing fast; sustained multi-hour office work is still mostly out of reach.

That’s the week. Stay sharp.