OpenAI released gpt-realtime-2.1 and gpt-realtime-2.1-mini on July 6, updating its Realtime API lineup for voice and multimodal agent development (MarkTechPost). The two models cut p95 voice latency by at least 25% compared to their predecessors through improved caching (OpenAI). The full-size model adds improvements to alphanumeric recognition, silence and noise handling, and turn-taking behavior, and now supports configurable reasoning effort for voice-agent workflows that require tool use and structured instruction following (MarkTechPost). Pricing for gpt-realtime-2.1 is $4.00 per million text input tokens, $24.00 per million text output, $32.00 per million audio input, and $64.00 per million audio output; gpt-realtime-2.1-mini comes in at $0.60 text input, $2.40 text output, $10.00 audio input, and $20.00 audio output per million tokens, aimed at cost-sensitive applications where lower reasoning depth is acceptable (MarkTechPost).
Anthropic’s Claude Fable 5 model moved off flat-rate subscription plans on July 7, shifting to metered usage credits for all tiers (Anthropic). Pro, Max, Team, and select Enterprise plans had included up to 50% of weekly usage limits against Fable 5 at no additional charge since the model returned to subscriptions earlier this month; that allowance ended today (TechTimes). The model now requires enabled usage credits, priced at $10 per million input tokens and $50 per million output tokens - double the rate of Claude Opus 4.8 at $5 and $25 respectively (BleepingComputer). Anthropic described the pullback as a temporary capacity measure driven by higher-than-expected demand, and said Fable 5 will return to standard subscription tiers once sufficient infrastructure is available (Anthropic).
Meta Research published SWE-Together, a multi-turn coding agent benchmark built from 11,260 recorded real user-agent sessions and distilled into 109 repository-level tasks (arXiv). Unlike SWE-Bench variants that score a single-turn patch against a closed test suite, SWE-Together requires agents to respond to live user clarifications and corrections as the session unfolds, using an anchored LLM simulator that releases conditional feedback when triggering conditions arise in the agent’s trajectory (arXiv). The benchmark introduces two evaluation dimensions beyond final code correctness: User Correction, measuring how much steering the simulated user must apply to keep the agent on track, and Intent Coverage, measuring whether the agent consistently surfaces user goals across evaluation runs (arXiv). The dataset and evaluation harness are released as open-source alongside the paper.