Cheaper models meet tougher agent tests
Agent economics just shifted. Three releases cut the cost of capable model work, while new training and evaluation updates offer a clearer way to judge whether agents can finish the job.
In today’s edition
- Claude Opus 5.5 lowers costs and speeds output.
- OpenAI cuts GPT-6 Sol and Luna API prices.
- OpenRouter launches half-price batch inference.
- Perplexity reports fewer tool-call failures after post-training.
- SWE-Bench Pro V2 refreshes its public benchmark.
Claude Opus 5.5 cuts typical costs by 40%
Anthropic’s new Opus model targets bigger coding and agent jobs with lower runtime costs. Claude Opus 5.5 is priced at $4 per million input tokens and $20 per million output tokens, and Anthropic says typical workloads cost 40% less than Opus 5.
- Cache reads cost $0.20 per million tokens, which Anthropic says is 60% less than Opus 5.
- Anthropic says Opus 5.5 generates output more than 30% faster than Opus 5.
- An early tester used it to audit and fix a 200,000-line codebase in under three hours, compared with more than 20 hours for Opus 5.
- Anthropic says the model uses fewer tokens per task as well as lower per-token pricing.
What it means Reprice long-running Claude agents, because lower token use, cheaper cache reads, and faster output can materially reduce the cost of codebase-scale work.
Source: Anthropic
OpenAI adds two cheaper GPT-6 model tiers
GPT-6 Sol and GPT-6 Luna bring much of GPT-6 Astra’s underlying advances to faster, more affordable models for work at scale, according to OpenAI. Their API prices are 50% lower than GPT-5.6 promotional pricing.
- OpenAI says both models build on the advances behind GPT-6 Astra.
- The company attributes the lower prices to more efficient caching and inference.
- Sol and Luna are available in the API.
- OpenAI says the new pricing gives teams more flexibility and room to iterate.
What it means Test Sol and Luna against your current production model, because a 50% API price cut can change which features and agent runs are economical to ship.
Source: OpenAI
OpenRouter makes asynchronous inference half-price
The OpenRouter Batch API is live for work that can wait. It cuts per-token pricing in half on most models, with results returned within 24 hours.
- The API supports Chat Completions, Responses, Messages, and embeddings across more than 70 models.
- OpenRouter reports that, during its two-week beta, more than 230,000 batches finished with a seven-minute median completion time and a 90th percentile under one hour.
- You submit one JSON body, receive a batch ID, then poll for completed results keyed by your custom IDs.
- Images and files must be public URLs, while audio, video, and OpenRouter’s web plugin are unavailable in batch requests.
What it means Move evaluations, enrichment, classification, and backfills to the batch path when results can wait, cutting eligible model spend by roughly half.
Source: OpenRouter
Perplexity trains agents on corrected mistakes
Perplexity post-trained a Computer model with hint-guided self-distillation to learn from its own errors. The company says a later checkpoint reduced tool-call failures by 21.2% relative to an earlier one in a live A/B test.
- Perplexity combines rejection-sampling fine-tuning with hint-guided self-distillation.
- Its approach preserves useful steps while correcting avoidable mistakes such as bad tool calls, including mistakes in sessions that eventually succeeded.
- The annotation process traces feedback to the decision responsible for an error and checks that corrective hints rely only on information available before the mistake.
- In a separate live test, tool-call failures fell from 2.24% to 1.77% between trained versions, without hints at inference time.
What it means When training agents from production traces, keep useful partial work and tie corrective feedback to the decision that caused the failure instead of labeling whole sessions good or bad.
Source: Perplexity
SWE-Bench Pro V2 refreshes coding-agent grading
SWE-Bench Pro V2 is a benchmark refresh, not a new coding agent. Scale says the updated public split uses a modified benchmark and locked evaluation protocol, with more of the grading evidence available for inspection.
- The release includes probe scripts, re-grade agents, per-task grades in both modes, and agent trajectories.
- SWE-Bench Pro evaluates challenging, long-horizon software-engineering tasks in public open-source repositories.
- The benchmark contains 1,865 tasks across 41 professional repositories.
- Scale says a task counts as resolved only when its submitted patch satisfies two strict conditions in the evaluation environment.
What it means When comparing coding agents, favor evaluations that publish task-level grades and trajectories, because a leaderboard score alone can hide how a result was produced.
Source: Scale AI Labs