[AINews] Opus 5.5 is good at explainer videos

What it claims

A one-day digest (AI News for 9/24–9/25/2026). Its headline is that Claude Opus 5.5 can autonomously produce explainer videos written entirely as code, with no diffusion model, at single-digit-dollar API cost. The summary names the releases. What it leaves out are the operating details: the reasoning-effort trap, judge and reranker cost and accuracy tradeoffs, licences, and several practical setup notes.

Evidence

Opus 5.5

• Leads SimpleBench at 88.4%.
• Ranked Anthropic's best vision model. It beats Fable 5 and GPT-6 Sol but trails GPT-6 Astra, at about 60% lower cost than Fable 5.1.
• On Terminal-Bench-Science it scores 24% at low effort, 62% at xhigh, and drops to 59% at max. Astra and Opus 5.5 lead Fable 5.1 by about 20 points. The best model from any other lab is Qwen3.8 Max at 12%.

Code-rendered video (Reddit reports)

• One-shot Claude Code runs produced 30–60s pure-JavaScript canvas animations with script, assets, TTS, music, and MP4 render.
• Costs reported: about $20 of Opus usage plus $3.21 of OpenRouter in about 1h20m; about $4 in 1.5–2h; one user needed "minor corrections."
• An interactive 3D island game (TideWater) took about 8 hours and about $1,874 in tokens, or 59% of a Max20x weekly allowance.

Jev (TypeSafe)

• $0.044 per 1K judgments at 152ms median latency, about 277× cheaper than GPT-6.
• Within 3 points on RewardBench and HaluEval, but 14.5 points behind on JudgeBench.
• A cascade that sends low-confidence calls to GPT-6 Astra keeps 99% of accuracy at 57% of cost.
• Ramp matched GPT-5.6 Luna reranking with 10× lower tail latency (300ms) at 3× lower cost.
• Counter-evidence from Reddit: on Banking77, BGE-small plus logistic regression scored 93.3% vs Jev's 83.2%, at about 9ms local inference.

Other results

• CLM, an open head on Qwen3-8B: 4–13× lower latency than Jev in agent benchmarks. As a verifier it scores 87.6% on Terminal-Bench 2.1 and 81.6% on DeepSWE (Jev is about 71% on DeepSWE). It is weaker zero-shot: BFCL v4 95.2% vs 99.2%.
• Harness-Zero: harness-free macro task success rises from 23.3% to 44.3%, beating the base model with the harness (41.7%).
• XYEval: one confident, misleading user hint cuts agent scores by up to 46.7% relative. Agents often disagree with the hint in their reasoning, then follow it anyway.
• Liquid DSpark (drafter for LFM2.5-VL-3B): speedups of 3.13× with MLX on M5 Max, 2.14× with llama.cpp on M3 Ultra, and 2.66× with SGLang on H100. Output quality is reported unchanged.
• Perplexity Photon: p99 fell from about 800ms to about 65ms on about 20% fewer machines. The Fast Search API runs at 160ms p50 / 230ms p95.

How to use it

• Opus 5.5: use xhigh, not max. Max imposes a minimum reasoning budget and scored lower.
• Weaviate 1.39 MMR: set balance explicitly. The default of 0.0 means pure diversity.
• CLM on limited GPUs: serve a Qwen3-8B GGUF (Q4_K_M, Q5_K_M or Q8_0) with: `` llama-server --embedding --pooling last `` The heads were trained on last-token representations, and mean pooling degrades scores. The heads are about 75 MB and fine-tunable.
• GGUF in Transformers: load directly with AutoModelForCausalLM.from_pretrained(..., gguf_file=...). On an M2 Max, throughput is close to llama.cpp (Qwen3.5-4B Q4_K_M: 70.4 vs 71.8 tok/s). Commenters expect LoRA training over GGUF via Unsloth or Axolotl, including MoE models, which bitsandbytes lacks.
• Xiaomi MiMo-V2.6-Pro:
• MIT licence, 1M context, AA index 46 (vs GPT-5.6 Sol at 47).
• Costs $0.13 vs $1.99 per task.
• RL code and training environments are also released.
• UkisAI Swift 1.5 27B: 58.5% fewer thinking tokens, +0.35% score vs base. Available in GGUF, NVFP4, MLX and W4A16.
• Gemini 3.8 Flash: free in Cline.

Limits

• This is a digest of tweets and Reddit threads. Most numbers are second-hand claims, not verified results.
• Jev: its "0% hallucination" claim guarantees only schema-valid output, not correctness. Critics say the right baselines are classifiers and rerankers, not LLM JSON generation.
• CLM:
• Calibrated context is only 2K–8K, vs Jev's 64K.
• Its probabilities are normalized over the supplied candidates only. A commenter suggests adding a "none-of-the-above" candidate.
• It is weaker at broad zero-shot use.
• MiMo: its RL gains reportedly don't generalize to harder math evals, and users report overthinking.
• Gemini 3.8 Flash on ARC-AGI v3: 10.4% with the standard harness vs 35% with the provider harness. Check which harness any figure you rely on used.
• Claude's ART enzyme finding: the short RNAs are confirmed, but the system's function and any editing utility are unknown.
• The Opus video results could not be independently inspected (Reddit returned 403).