Verified 2026-09-26

By Crypto Nite1 min read

This Week in Coding Intelligence: Opus 5.5 Meets GPT-6 Sol

On 22 Sept 2026, Anthropic launched Claude Opus 5.5 at / per M tokens—heaper than Opus 5, cache reads (f)—while OpenAI shipped GPT-6 Sol/Luna into…

Contents

    The Week in One Paragraph

    Both frontier labs moved the cost floor in the same 24 hours. Anthropic shipped Claude Opus 5.5 on 22 September 2026 — 40% cheaper to run than Opus 5, with cache reads down 60% — while OpenAI pushed GPT-6 Sol and GPT-6 Luna into Codex and the API and halved GPT-5.6 promotional pricing. OpenAI also rebuilt GPT-6 prompt caching around a 30-minute reuse window and discounts of up to 90% on cached input. On 25 September Anthropic opened a Claude plugin directory and confirmed MCP 2.0 support, and Claude Code’s 2.1.280–2.1.283 releases added AGENTS.md, admin model governance, and a prompt-audit command. The through-line: agent capability held roughly steady while the cost of running agents fell sharply.

    Anthropic ships Claude Opus 5.5 — the agentic-coding cost curve bends

    Anthropic’s Claude Opus 5.5 landed on 22 September 2026 as the first model in the Claude 5.5 family, priced at $4 per million input tokens and $20 per million output tokens — 40% cheaper to run than Opus 5, with cache reads at $0.20 per million and output generation more than 30% faster.

    • Opus 5.5 performs “at the level of Claude Fable 5.1 on most work” and costs 40% less to run than Opus 5, with cache reads 60% cheaper at $0.20 per million tokens, per Anthropic’s announcement.
    • Anthropic reports it is the strongest model to date on its automated behavioral audit, less likely than recent models to take hard-to-reverse actions, and more resistant to prompt injection than Opus 5, following pre-release external testing by Frontier Design and METR with a published system card.
    • Reported capability datapoints include one tester completing a 680,000-line code migration in less than a day, and an internal web-app load-time task succeeding 39 of 40 times where Opus 5 made smaller changes that also altered app behavior.
    • It ships with “preserved thinking,” the anti-distillation safeguard from Fable 5.1, applying to API accounts created on or after 31 August 2026; thinking can no longer be switched off, EU AI Act watermarking applies, and the platform model id is claude-opus-5-5, available on AWS, Google Cloud and Azure.

    What This Changes: Long-horizon migration and refactor work is now the default target for Opus 5.5, because fewer unintended behavior changes are the expensive failure mode in those jobs. Keep external approval gates on irreversible actions regardless of improved behavioral audit scores, and budget cache reads explicitly — at $0.20 per million tokens, a stable prefix is a cost decision, not just a latency one.

    OpenAI ships GPT-6 Sol and GPT-6 Luna into Codex + API and cuts prices 50%

    OpenAI shipped GPT-6 Sol and GPT-6 Luna into Codex and the API on 22 September 2026 and halved prices: Sol at $2/$10 per million input/output tokens and Luna at $0.10/$0.50. Sol scores 33.2% on AutomationBench 1.0.6 at xhigh effort for $0.27 per task.

    • GPT-6 Sol moved from $4 to $2 per million input tokens and $20 to $10 per million output tokens — 50% cheaper than GPT-5.6 promotional pricing. GPT-6 Luna, the lowest-cost GPT-6 tier, went from $0.20 to $0.10 input and $1.20 to $0.50 output, per OpenAI’s launch post.
    • Both were trained with similar methods to GPT-6 Astra, carrying its advances in professional work, factuality, coding, computer use and alignment to faster, cheaper models; GPT-6 Astra remains the top model. Third-party Codex timelines report both added to Codex and the API as gpt-6-sol and gpt-6-luna for Plus, Pro, Business, Enterprise and Edu.
    • On AutomationBench 1.0.6 (47 tools across sales, marketing, operations, support, finance and HR), GPT-6 Sol at xhigh effort scores 33.2% at $0.27 per task, while Claude Opus 5 at max effort scores 26.9% at 11.1× that cost and GPT-6 Astra at low effort scores 30.3% at 3.9× the cost. At high effort, GPT-6 Luna improves on its predecessor by 5.4 percentage points at 58% lower cost per task.
    • On Agents’ Last Exam V1, GPT-6 Sol at max effort scores 56.4%, above Claude Opus 5’s highest evaluation score, at 60% lower cost per task.

    What This Changes: For automation agents where cost-per-task dominates the routing decision, GPT-6 Sol is now the default candidate and Luna is the cheap tier worth measuring. Run your own benchmark against existing Opus 5 flows before switching, because the published numbers are vendor- and benchmark-specific, and reserve GPT-6 Astra for the hardest tasks where its higher cost per task buys accuracy you cannot get elsewhere.

    OpenAI rebuilds GPT-6 prompt caching — the harness-layer economics story

    OpenAI rebuilt GPT-6 prompt caching around higher hit rates and a 30-minute reuse window, discounting cached input tokens by up to 90%. A new dashboard and a prompt_cache_diagnostics object explain misses, while explicit breakpoints, mid-conversation configuration updates and cache prewarming move processing out of user wait time.

    • The new default caching system delivers higher hit rates, with discounts for eligible shared prefixes reused within a 30-minute window and up to 90% off cached input tokens, per OpenAI’s caching post.
    • A new Prompt Caching Dashboard shows cache hit rate over time plus an input-composition chart comparing cached and uncached tokens, and a diagnostics tool returns a prompt_cache_diagnostics object with the fields type, reason, comparison_reusable_tokens and cache_missed_tokens, documented in the API caching guide.
    • Harness-level controls now include explicit cache breakpoints, the ability to change reasoning effort mid-conversation without breaking cache via a configuration_update appended while request-level effort is unchanged, append-only tool updates (allowed_tools, tool_choice none instead of removing definitions), and cache prewarming.
    • GitHub Copilot (Mario Rodriguez, CPO) reports cutting the share of prompt tokens requiring fresh processing by more than 50% across billions of requests versus its previous baseline.

    What This Changes: Agent harnesses should be redesigned around stable prefixes and append-only tool definitions rather than mutating tool lists in place. Emit configuration_update instead of rewriting request-level reasoning effort, prewarm cache outside user-facing latency, and wire prompt_cache_diagnostics into your traces so hit-rate regressions surface as a metric rather than a surprise invoice.

    Anthropic opens the Claude plugin directory and ships MCP 2.0 support

    Anthropic opened a Claude plugin directory on 25 September 2026 and confirmed support for the latest MCP specification, dated 2026-07-28 and known as MCP 2.0, with a stateless core plus MCP Apps and Enterprise Managed Auth. Plugins are now pitched as the main way to build third-party Claude extensions.

    • Plugins are now “the main way to build third-party extensions for Claude,” with a directory submission portal open to developers on paid Claude plans, per Anthropic’s plugin announcement.
    • Two submission paths exist: a single MCP connector pointing at a remote MCP server, or a plugin bundle combining MCP servers and skills hosted on GitHub. In Claude Code, plugins can also include LSPs, commands, hooks and agents.
    • The portal performs auto-validation and a safety scan at submission, shows review status with recommended changes, supports publish-when-you-choose, and offers post-launch analytics covering installs by product surface and version, listing views, and which searches led users to the listing.
    • Claude supports the MCP specification dated 2026-07-28 with a stateless core, plus two extensions: MCP Apps for interactive UI inside chat and Enterprise Managed Auth for zero-touch OAuth, as described in the MCP blog. Skills and MCP connectors remain building blocks and stay listed in the directory.

    What This Changes: If you maintain internal MCP servers, packaging them as directory plugins is now the distribution and telemetry path — install counts by product surface and search referrers are adoption data you would otherwise instrument yourself. Adopt the stateless MCP 2.0 core early to avoid proxy-state failures, and prefer Enterprise Managed Auth over per-user OAuth flows for enterprise tenants.

    Claude Code 2.1.280 → 2.1.283: AGENTS.md, model governance, prompt-audit

    Claude Code’s 19–25 September releases added AGENTS.md project instructions, admin model governance, and a prompt-audit command. Version 2.1.280 introduced AGENTS.md, 2.1.281 extended it to cloud providers and gateways, and 2.1.283 added gateway hint headers, model allow/deny settings, prompt auditing and MCP resilience fixes.

    • 2.1.280 added AGENTS.md support: in a project with no CLAUDE.md, Claude Code reads AGENTS.md instead, configurable under “Project instructions” in /config. 2.1.281 extended that support to Amazon Bedrock, Google Vertex AI, Microsoft Foundry, LLM gateways, and sessions with telemetry disabled, per the official changelog.
    • 2.1.283, first seen 25 September, added the x-claude-code-prompt-id gateway hint header (opt in with CLAUDE_CODE_GATEWAY_HINT_HEADERS=1) so LLM gateways can group requests serving one user prompt, plus availableModelsMatch: "exact" and deniedModels managed settings so admins can keep new model versions blocked until explicitly listed.
    • The same release added /doctor prompt-audit (also /checkup prompt-audit) to audit CLAUDE.md files, skills, agents and commands for prompting patterns written for older models, and added MCP tool, WebFetch and WebSearch outputs to the OpenTelemetry tool.output span event when OTEL_LOG_TOOL_CONTENT=1.
    • Resilience fixes relevant to harness builders: a brief HTTP 404 from a stateless remote MCP server, such as a proxy mid-redeploy, no longer leaves that server unusable for the whole session while still shown as connected; stdio MCP servers are no longer left running when the session ends while they are still starting; and MCP progress notifications are no longer discarded once a long-running tool call moves to the background.

    What This Changes: Adopt AGENTS.md as a fallback project-instruction standard so repositories work across tools without duplicating guidance, and add prompt-audit to your model-upgrade checklist — stale prompting patterns are a silent quality regression when a new model ships. Use deniedModels with availableModelsMatch: "exact" to keep unvetted model versions out of managed fleets, and turn on the prompt-id header if you route Claude Code through a gateway and need per-prompt request grouping.

    The Bottom Line

    The two announcements that matter most this week are pricing moves, not capability moves. Anthropic cut Opus 5.5 to $4/$20 per million tokens with 60%-cheaper cache reads (announcement), and OpenAI halved GPT-6 Sol and Luna prices (launch post) while publishing cost-per-task numbers that put Sol ahead of Claude Opus 5 on AutomationBench 1.0.6 at a fraction of the cost. Meanwhile OpenAI’s prompt caching rebuild pushed cached-input discounts to 90% for reusable prefixes, and GitHub Copilot’s reported >50% reduction in fresh prompt tokens shows the harness layer is where those discounts actually get captured.

    The practical read for teams running coding agents: reprice your routing now. Re-run your own evaluations against Opus 5.5, GPT-6 Sol and GPT-6 Luna rather than trusting published benchmark tables, because the brief’s numbers are vendor-reported and benchmark-specific. Then spend the saved budget on harness work — stable prefixes, configuration_update, append-only tool definitions, and prewarming — since that is what converts list prices into real bill reductions.

    The second-order story is governance catching up with distribution. Anthropic’s plugin directory adds validation, review status and post-launch install analytics; Claude Code 2.1.283 adds model allow/deny lists, prompt-id headers and a prompt-audit command. If you ship agents to other people’s repositories or route them through a gateway, those controls arrive at roughly the right time — adoption is about to be measurable, and so is drift.

    How This Briefing Was Researched

    Desk research on 26 September 2026 over the trailing Saturday-to-Friday week (19–25 September 2026). Sources are primary: vendor announcement pages from Anthropic, OpenAI and Claude, the official Model Context Protocol specification and blog, and the official Claude Code CHANGELOG.md. Trackers such as the Evertune model-release dashboard, llm-stats and Releasebot were used only to find items; every date, price and capability claim here traces to the vendor’s own page. No hands-on testing was performed — no model was run and no API was called. Every external link cited above returned HTTP 200 when fetched.

    FAQ

    Should I switch from Claude Opus 5 to GPT-6 Sol for coding agents?

    Not on benchmark tables alone. On AutomationBench 1.0.6, GPT-6 Sol at xhigh effort posted 33.2% at $0.27 per task against Claude Opus 5 at max effort at 26.9% at 11.1× that cost, and Sol scored 56.4% on Agents’ Last Exam V1. Run your own evaluations against your actual task mix before switching.

    How do I avoid breaking GPT-6 prompt cache when changing reasoning effort or tools?

    Use the mechanisms OpenAI added rather than rewriting requests. Change reasoning effort mid-conversation through a configuration_update appended while request-level effort stays unchanged, and update tools in an append-only way — pass allowed_tools and set tool_choice to none instead of removing tool definitions. Then watch prompt_cache_diagnostics for cache_missed_tokens to confirm the fix.

    Do I need to adopt MCP 2.0 plugins now?

    No deadline exists, but the migration is low-risk and compounding. Claude supports the 2026-07-28 MCP spec with a stateless core plus MCP Apps and Enterprise Managed Auth, while skills and MCP connectors remain supported building blocks. If you maintain internal MCP servers, the stateless core removes proxy-state failure modes you are probably already handling by hand.

    📖 Related Reads

    • ToolBrain — tool reviews, LLM comparisons, and AI workflow guides
    • Hermes Tutorials — Hermes Agent setup, configuration, and advanced workflows
    • NoCode Insider — AI workflow automation with no-code tools, agents, and APIs

    Cross-links automatically generated from CodeIntel Log.

    More in weekly →