Agent power is no longer the bottleneck

Builder Brief

→ GPT-5.6: OpenAI added Sol, Terra, and Luna model sizes, Programmatic Tool Calling, beta multi-agent support, explicit cache breakpoints, and a 30-minute minimum cache life. Source

→ Fable 5: Anthropic restored access on July 1 after export controls were lifted, and now routes blocked Fable 5 requests to Opus 4.8. Source

→ Sonnet 5: Intro pricing is $2 per million input tokens and $10 per million output tokens through August 31, then $3 and $15. Source

→ Voice and work agents: GPT-Live powers a new ChatGPT Voice experience, while ChatGPT Work is positioned as an agent that acts across apps and files. Voice · Work

→ Meta Muse Spark 1.1: Meta is previewing a Model API for a multimodal agent model with 1M-token context, computer use, coding, parallel subagents, and MCP-style tool use. Source

→ Enterprise platforms: Google and Microsoft are converging on the same agent pattern: build, scale, govern, evaluate, and optimize. Google · Microsoft

→ Apple v. OpenAI: Apple sued OpenAI and two former Apple employees on July 10, alleging systematic theft of hardware trade secrets tied to OpenAI's device push. Not a builder how-to, but a reminder that cross-lab hiring and IP boundaries are now litigation risk, not just HR policy. Source

New from Best-AI.org

Daily AI News is now live on the App Store

Our new iOS app, Daily AI News (Best-AI.org), is available on the Apple App Store. It brings our AI news, podcasts, article audio, bookmarks, and community voting into a native mobile experience for iPhone and iPad. Android is coming soon.

Use it when you want to follow AI news away from your desk: read in dark mode, listen to article audio, save pieces for later, and keep your Best-AI.org profile synced across devices.

This Month's Briefing

OpenAI

GPT-5.6 makes agent design more programmable

OpenAI introduced GPT-5.6 on July 9 with three API model sizes: Sol, Terra, and Luna. The release adds Programmatic Tool Calling in the Responses API, beta multi-agent support, more predictable prompt caching, explicit cache breakpoints, and a 30-minute minimum cache life.

What to build now: Add task-level evals before switching defaults. Measure cost, latency, tool-call count, retry rate, quality, and final task success on your own workflows. The biggest model should not automatically become the default for every request.

Source: OpenAI, GPT-5.6

Voice and Work Agents

GPT-Live and ChatGPT Work move agents into normal workflows

OpenAI introduced GPT-Live on July 8 as a new generation of voice models powering ChatGPT Voice. One day later, OpenAI positioned ChatGPT Work as an agent that can act across apps and files, stay with a project for hours, and turn a goal into finished work.

What to build now: If your product has a voice, support, meeting, or workflow surface, design for interruption, confirmation, handoff, and recovery. A useful voice agent is not just a speech model. It is a turn-taking system with memory boundaries and clear escalation paths.

Sources: GPT-Live · ChatGPT Work

Meta

Muse Spark 1.1 brings another agent model into the API market

Meta introduced Muse Spark 1.1 on July 9 and launched a public preview of the Meta Model API. Meta describes the model as a multimodal reasoning model for agentic tasks, with gains in tool use, computer use, coding, multimodal understanding, and a 1 million token context window.

The more interesting product signal is orchestration. Meta says Muse Spark 1.1 can act as a main agent, delegate execution across parallel subagents, manage long context, and use tools such as MCP servers and custom skills.

What to build now: Keep your agent architecture provider-portable. Define tool schemas, permissions, evals, context compaction, and audit logs outside a single model vendor so you can test OpenAI, Anthropic, Meta, and cloud-hosted options without rewriting the product.

Source: Meta AI, Muse Spark 1.1

Anthropic

Fable 5 returned, but the lesson is fallback routing

Anthropic said access to Claude Fable 5 and Mythos 5 was restored on July 1 after export controls were lifted on June 30. Fable 5 returned globally on Claude Platform, Claude.ai, Claude Code, and Claude Cowork. Mythos 5 was restored for approved US organizations while Anthropic continued coordinating broader Glasswing access.

The production detail matters: Anthropic says its new classifier blocks the described bypass technique in more than 99% of cases, but it may also create more false positives for benign coding and debugging. Blocked Fable 5 requests are routed to Opus 4.8.

Anthropic followed up on July 2 with a proposed Cyber Jailbreak Severity framework and a HackerOne program for potential Fable 5 cyber jailbreaks. The framework grades jailbreaks by capability gain, breadth, ease of weaponization, and discoverability.

The pricing story is also moving. Anthropic originally planned to move Fable 5 off included weekly limits on July 7, then pushed that deadline to July 12, and pushed it again to July 19 after user pushback. Until then, Pro, Max, and Team plans can still use Fable 5 for up to 50% of their weekly limit at no extra cost. After that window, further use draws from prepaid usage credits at a confirmed $10 per million input tokens and $50 per million output tokens, the highest published rate for a generally available Claude model.

What to build now: Treat safety blocks and jailbreak reports as normal runtime events. Log the request class, model, fallback model, user-visible message, cost impact, severity estimate, and whether the final task succeeded. If you meter Fable 5 for customers, do not hard-code the July 19 cutover into your billing logic: this deadline has already moved twice.

Sources: Redeploying Fable 5 · Cyber safeguards framework · BleepingComputer, access extension

Model Market

Claude Sonnet 5 changes the middle tier

Anthropic introduced Claude Sonnet 5 as the default model for Free and Pro plans, and made it available to Max, Team, Enterprise, Claude Code, and the Claude Platform. Introductory API pricing is $2 per million input tokens and $10 per million output tokens through August 31, 2026. Standard pricing after that is $3 and $15.

What to build now: Create three routing tiers: fast default, high-effort default, and premium specialist. Then test Sonnet 5 against your existing mid-tier model on real support, coding, research, and operations tasks.

Source: Anthropic, Claude Sonnet 5

Enterprise Agents

Agent platforms are becoming governance platforms

Google Cloud's AI roundup highlights Gemini Enterprise Agent Platform as a place to build, scale, govern, and optimize agents. Microsoft's Work IQ APIs, generally available since June 16, point in the same direction: give agents a shared context and workflow layer instead of wiring each one to Microsoft 365 by hand.

The coalition politics matter as much as the features. In June, Google, Microsoft, Salesforce, and eight other vendors, including ServiceNow, Snowflake, Databricks, and GitHub, published the Agentic Resource Discovery (ARD) specification, an open standard for agents to find and verify tools across company software. ARD does not replace Anthropic's Model Context Protocol or Google's own Agent2Agent protocol; it sits above them as a discovery layer. Read it as a signal that the largest platform vendors do not want one AI lab controlling how agents connect to enterprise software.

What to build now: Write your agent operating model. Include who can create agents, which systems they can access, how credentials are stored, when approvals are required, and how failed actions are reviewed. Do not bind your architecture to one discovery or tool-calling standard; MCP, A2A, and ARD are all active at once.

Sources: Google Cloud · Microsoft, Work IQ APIs · Google, Agentic Resource Discovery

Builder Playbook

Use this model-routing matrix before you change defaults

The new model launches are useful only if your product can route work intentionally. Start with this lightweight matrix, then replace the placeholders with your real latency, quality, and cost data.

The Number

73.5%

OpenAI reports GPT-5.6 scored 73.5% on ExploitBench 2, compared with 47.9% for GPT-5.5 at a comparable output-token budget. The builder implication is simple: if your tool writes, reviews, scans, or executes code, define boundaries for vulnerability analysis, exploit generation, logs, and human review.

Editor's Pick

Claude Sonnet 5

Sonnet 5 is a serious candidate for the middle tier of production agent workflows. Try it where you currently use a cheap model with many retries, or a premium model for tasks that do not always justify premium cost. Compare it against GPT-5.6 Luna and Muse Spark 1.1 when you evaluate routine agent tasks.

One Thing To Try

Run a portability and fallback drill

Pick one AI workflow and simulate five events: primary model unavailable, safety classifier block, cost budget exceeded, tool call failure halfway through, and a switch to a different provider.

Your product should preserve the task, explain the degraded state, switch to a configured fallback when appropriate, and record why the fallback happened.

You Asked, We Answered

Should we move every AI feature to the newest frontier model?

No. Start with a routing table, not a migration announcement. Use the newest model for high-value tasks where its quality changes the outcome. Use mid-tier models for routine workflows. Use smaller or cached-context flows where latency and cost matter more than maximum reasoning.

If You Read One Thing

Anthropic's cyber safeguards and jailbreak framework

It is the most useful operational read in this issue. It separates prohibited, high-risk dual use, low-risk dual use, and benign cybersecurity requests, then proposes a Cyber Jailbreak Severity scale. That is directly useful if you are designing safety triage for an AI tool.

Cost Model

Stop tracking cost per token. Track cost per useful outcome.

A cheaper model can be more expensive if it retries, fails, or needs human repair. A stronger model can be cheaper if it completes the workflow with fewer tool calls and less rework.

Ship this week

A 5-point control checklist for AI tool builders

1. Add a model-router config table with primary, fallback, low-cost, high-effort, blocked-request, and cross-provider behavior per task type.

2. Track cost per successful task, including retries, cached input, uncached input, tool calls, and failed runs.

3. Create a safety-block UX state that tells users what happened without exposing policy internals or sensitive details.

4. Add an agent action log with user, provider, model, tool, input class, output class, approval status, severity estimate, and fallback reason.

5. Write one eval set from real user tasks before adopting GPT-5.6, Sonnet 5, Muse Spark 1.1, or any other new default.

Developer Takeaway

The durable AI product layer is operational: routing, evals, controls, fallback behavior, and observability. Model quality still matters, but users remember whether the workflow completed, whether the answer was trustworthy, and whether your product recovered when something went wrong.

1. Route by task: match model size, effort, cost, and risk to the actual job.

2. Measure real outcomes: cost per successful task beats cost per token.

3. Treat blocks and fallbacks as product states: explain them clearly and log them cleanly.

4. Govern agents early: identity, permissions, credentials, approvals, and rollback are part of the feature.

Built an AI tool?

Submit it to Best-AI.org and put it in front of people actively comparing AI products.