One-Line Summary

OpenAI cuts everyday model costs and improves caching while new research shows how to stabilize 8‑bit reinforcement learning and keep models/agents up to date, and open-source stacks for serving and UI mature.

LLM & SOTA Models

OpenAI releases GPT-6 Sol and Luna with 50% lower API prices

OpenAI is releasing two models meant to get everyday work done for less money by halving token prices and making reused context cheaper to read. GPT‑6 Sol is priced at $2 per million input tokens and $10 per million output tokens (down from $4/$20 for GPT‑5.6 Sol); GPT‑6 Luna costs $0.10/$0.50 (down from $0.20/$1.20). They are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, with Luna also in the desktop app for Free and Go users; in the OpenAI application programming interface (API) they appear as “gpt‑6‑sol” and “gpt‑6‑luna,” and they are not yet available in Chat (as of Sep 22, 2026). ¹

On end‑to‑end workflow benchmarks, OpenAI reports that GPT‑6 Sol at extra‑high effort scores 33.2% on AutomationBench at $0.27 per task and outperforms Claude Opus 5 at a fraction of the cost; on Agents’ Last Exam, Sol reaches 56.4% at 60% lower task cost than Opus 5 in OpenAI’s setup. For coding, Sol scores 68.8% on DeepSWE 1.1 (vs. 69.9% for Claude Fable 5 at higher effort) with roughly 80% lower task cost, and Luna reaches 66.6% with much lower costs than the compared Anthropic configs. OpenAI also says Sol makes about half as many factual mistakes as GPT‑5.6 Sol in its internal evaluation, approaching Astra‑level reliability at lower cost. ¹

OpenAI is also changing the economics of long conversations and agents by improving prompt caching. The company says GPT‑6 delivers higher default cache hit rates and a 90% discount on cached input‑token reads; developers can change reasoning effort or enable/disable tools without breaking cache, and a diagnostics dashboard explains missed cache opportunities. GitHub reports that these changes cut fresh prompt tokens by more than 50% across billions of requests to OpenAI models. ¹

In competitive context, VentureBeat notes OpenAI positions Sol and Luna’s prices as permanent and emphasizes cost per completed task. Anthropic’s new Opus 5.5 costs $4/$20 per million tokens, Claude Sonnet 5 sits at $2/$10, and Google’s Gemini 3.8 Flash is $0.75/$3.75 through Dec 31, 2026 before increasing to $1.50/$7.50. VentureBeat also points out there is not yet a same‑harness public result comparing Sol with Opus 5.5 on cost per successful task. ²

Open Source & Repos

SGLang speeds up serving for large and multimodal models

SGLang is a high‑performance serving framework for large language models and multimodal models, aimed at low‑latency, high‑throughput deployments. The v0.5.20 release highlights active development with 713 pull requests from 237 contributors and adds support for additional models alongside updated docs and cookbook examples. ³

For teams building agents and assistants, a serving layer like SGLang helps batch requests efficiently and keep response times predictable as usage scales, making it easier to run production workloads on shared GPUs. ³

CopilotKit updates its agent-native UI stack for web and chat

CopilotKit is a frontend stack for agent‑native applications and generative UI across React, Angular, Vue, React Native, and channels like Slack and Microsoft Teams. The v1.73.3 update adds a smoother onboarding step (copy a hosted prompt link instead of a command‑line step) and fixes inspector accuracy for conversation results and failures. ⁴

It provides shared state, human‑in‑the‑loop workflows, and the AG‑UI protocol to connect any agent framework to common channels, helping product teams ship consistent, suggestion‑driven interfaces without rebuilding UI primitives from scratch. ⁴

Research Papers

Stabilizing full‑pipeline FP8 RL with calibrated clipping

This paper tackles unstable training when using 8‑bit floating point (FP8) across the full reinforcement learning (RL) pipeline for large language models (LLMs). The authors trace mid‑training entropy spikes and garbled outputs to compounded FP8 quantization noise that distorts the importance ratio, pushing negative‑advantage tokens outside the trust region and zeroing their gradients. They propose Calibrated Clipping, which aligns FP8 clipping bounds to bfloat16 (BF16) distributions by matching the lower‑bound quantile and rebalancing the upper bound. ⁵

Experiments across model scales from 8B to 32B and multiple FP8 scaling granularities show the method eliminates entropy surges and restores performance comparable to BF16 baselines, addressing instability beyond earlier train‑inference mismatch fixes. For teams training RL pipelines under strict hardware budgets, the result points to a path for maintaining FP8’s speedups without sacrificing stability. ⁵

Continual learning recipes for models and agents

Time‑Incremental Continued Pretraining of LLMs studies updating models only on post‑cutoff data and finds that six open‑weight models (OLMo2 1B/7B, Llama‑3.2 1B/3B, Llama‑3.1‑8B, Gemma‑3‑1B) acquire new knowledge without catastrophic forgetting. Macro‑averaged accuracy across thirteen downstream tasks stays within 0.01 of the base for every model; a curated 6B‑token slice matches a broader 40B slice; the optimal learning rate for knowledge acquisition differs by about an order of magnitude from the one for general capability; and low‑rank adaptation (LoRA) at sufficient rank matches full continued pretraining. Gains survive supervised fine‑tuning; direct preference optimization (DPO) interacts in a family‑dependent way. ⁶

ACLArena analyzes agent continual learning, comparing on‑policy multi‑teacher distillation, self‑distilled fine‑tuning, and model merging, and then proposes a recipe that combines offline replay on high‑quality trajectories with a routed network of LoRA experts, each specialized via RL, to preserve and extend capabilities across domains. ⁷

EvoPathBench introduces capability‑level, process‑oriented evaluation of self‑evolving agents, tracking when a capability emerges and whether subsequent updates strengthen, preserve, or weaken it. Across 1,800 streams built from hybrid real‑market trading data and generated trajectories, the authors find generalization gains can fade under distribution shift, forgetting concentrates in a minority of paths, and reliable rule adaptation remains unresolved—highlighting candidate evaluation and selection as key bottlenecks. ⁸

Community Pulse

Hacker News (931↑) — Reactions weigh big price cuts against hints of capability regression and worries about inconsistent token accounting across providers. ⁹

"The notable thing is that Luna regressed a bit on coding while dropping 60% in price.That’s a fair trade, for high-volume work Luna at that price is basically free, but it does show that newer doesn’t always mean better." — Hacker News ⁹

"Wow I cannot believe Luna is getting even cheaper. IMO this is the model that is going to change the world. Everyone said tokens were too expensive but these are getting close to free while still having fantastic performance." — Hacker News ⁹

Hacker News (197↑) — Interest in CopilotKit’s approach is tempered by calls for lower suggestion latency and clearer labeling around framework scope. ¹⁰

"Grow your OpenAI bill exponentially with this one easy trick! In all seriousness — looks nifty! Do you guys have any plans to bring suggestion latency down from what’s show in the demo?" — Hacker News ¹⁰

"It'd be nice if this was more clearly labeled as a React plugin. I saw "drop in replacement for " and thought it was a web component and only realized it was a react plugin after reading further. Not a huge deal, but since it is a plugin and not a stand-alone thing it might make sense to make that clear up front." — Hacker News 10

Why It Matters

Lower token prices and a 90% cached‑read discount shift the “cost per task” math for agents and long chats. With vendors also changing token accounting and cache policies, teams get the most clarity by measuring cost per successful workflow rather than just headline per‑token rates. ²

On the research side, stabilized full‑pipeline FP8 RL and time‑incremental updates indicate cheaper, steadier training and maintenance, while process‑level agent evaluation clarifies what to track (generalization, retention, rule adaptation) as serving and UI stacks like SGLang and CopilotKit make shipping those systems more practical. ⁵

This Week to Try

1. GPT‑6 Luna hands‑on: In the ChatGPT desktop app, pick “GPT‑6 Luna” and compare a document summarization or extraction task with your current default model. Details in the announcement. https://openai.com/index/introducing-gpt-6-sol-and-luna/
2. Run SGLang locally: Follow the GitHub quickstart to serve an open model and send a few prompts to see batching/latency behavior. https://github.com/sgl-project/sglang