Cheaper Frontier Models Arrive as AI Coding Agents Face New Security Flaws

This week brings good news for anyone paying for AI APIs and a warning for anyone running AI agents on their own machine. OpenAI and Anthropic cut prices on new flagship models, and SpaceXAI and GitHub shipped developer-focused releases [S6][S9][S10][S11]. At the same time, researchers and Meta disclosed security flaws in AI coding tools and a desktop agent that call for prompt updates [S13][S16].

GPT-6 Sol and Luna Halve API Costs

OpenAI expanded its GPT-6 family with Sol, priced at $2.00 per million input tokens and $10.00 per million output tokens, and Luna, priced at $0.10 input and $0.50 output. OpenAI says that is a 50% cut compared with the GPT-5.6 series. It also reports that factual error rates fell by roughly half and that accuracy improved on its internal coding and factuality evaluations [S6][S8].

For freelancers and solo developers paying per token, this matters directly. Coding, refactoring, and document tasks now cost about half as much. Luna's low price makes high-volume, lighter jobs cheaper to run. Keep in mind that the accuracy gains come from OpenAI's internal evaluations, so test the models on your own tasks before you rely on them [S6][S8].

Claude Opus 5.5 Targets Long, Cache-Heavy Workflows

Anthropic released Claude Opus 5.5 with a base API price 20% lower than before, at $4.00 per million input tokens and $20.00 per million output tokens. Prompt cache reads are 60% cheaper, at $0.20 per million tokens, and generation is reported to be 30% faster. On safety, Anthropic reports an 85% drop in containment bypass attempts in its testing [S9][S17].

The caching discount matters most if you reuse large contexts. Examples include long-running agent sessions, code reviews over the same repository, and software migrations. On these cache-heavy workloads, total API spending can fall by up to 40%, and faster output shortens waiting time in interactive work [S9][S17].

GitHub HydraFusion and Grok 4.7 Tune AI for Developers

GitHub launched Project HydraFusion as a research preview for Copilot CLI. Instead of sending every request to one model, it builds execution plans that route terminal requests across multiple models. GitHub reports estimated task costs 67% lower than Claude Opus 5 and a 4.9-point accuracy gain on TerminalBench 2.1. According to the coverage, Copilot users can turn it on with a single CLI flag, which makes it easy to test on multi-step terminal work. Because it is still a preview, expect changes [S11].

SpaceXAI's Grok 4.7 is aimed at software engineering and knowledge work and costs $2.00 per million input tokens and $6.00 per million output tokens. Its redesigned safeguards are meant to reduce accidental refusals on harmless developer tasks. On the HackerBench v0.3 test, it allowed 3.3% of dual-use prompts. For security professionals, the practical benefit is fewer interruptions during legitimate testing [S10].

Plugin4Shell: Update Your AI Coding Agents

Researchers disclosed Plugin4Shell, a zero-click remote code execution flaw affecting Claude Code, OpenAI Codex, GitHub Copilot, and Gemini CLI. The flaw stems from unverified git plugin replacements. Fixes have been verified in Claude Code v2.1.179 and Codex v0.146.0 [S13].

If you use these tools, confirm that you are on at least those versions. Fixes for the other affected agents were not verified in the reporting. Until they are, be cautious with untrusted plugins, especially in GitHub Copilot [S13].

Meta Hot-Fixes a Zero-Day in Its Muse Desktop Agent

Meta released an emergency fix for a zero-day flaw in Muse, its desktop AI agent. The flaw let local, unprivileged processes redirect dictation audio, inject untrusted prompts, and extract authentication tokens. If you run Muse, apply the update now to protect your dictation, personal context, and account credentials [S16].

What to watch next

Taken together, these stories show two trends moving at once. Model prices are falling and developer tooling is improving, which makes capable AI cheaper for individual professionals [S6][S9][S11]. But agents that access your files, plugins, and microphone are also becoming targets for attack. The practical response is to review which models you pay for and to keep every AI agent on your machine up to date [S13][S16].