Big 5 AI Vendor Roundup: Week of June 29, 2026

The Fable 5 saga ended this week, and the way it ended tells you where the industry is headed. The Commerce Department lifted its export controls on June 30, and Anthropic put Fable 5 back online globally on July 1, 19 days after it went dark. The same day the controls lifted, Anthropic shipped Claude Sonnet 5, its new everyday workhorse model. Two days earlier, OpenAI launched its GPT-5.6 family, but only to a small group of partners the US government approved first. Google’s Gemini 3.5 Pro missed its promised June launch and slipped to July, which leaves Google holding the only frontier model without government strings attached, mostly because it hasn’t shipped one lately. Microsoft made its in-house coding model generally available for Copilot Business and Enterprise. AWS had a quiet week while it works to get Fable 5 back onto Bedrock.

Put the two big stories side by side and the pattern becomes pretty clear. Anthropic got its model back by agreeing to give the government early access to future models, and OpenAI launched its new models through a government-approved preview list. Frontier releases now look like negotiated deployments, and that changes how IT leaders should plan around them.

Anthropic: Fable 5 is back, Sonnet 5 arrives, and Washington gets a seat at the table

• The export controls are gone, and Fable 5 is live again. Commerce Secretary Howard Lutnick lifted the controls June 30 in a letter to Anthropic co-founder Tom Brown, and Fable 5 returned globally July 1 on the Claude Platform, Claude.ai, Claude Code, and Claude Cowork. Pro, Max, Team, and select Enterprise plans get Fable 5 included for up to 50% of weekly usage limits through July 7, after which it moves to usage credits. Access on AWS, Google Cloud, and Microsoft Foundry is being reenabled “as quickly as possible.” Mythos 5 remains limited to roughly 100 approved US organizations, the carveout from June 26.
• What the two-week review actually found. Anthropic’s testing, done jointly with the government and Amazon, showed the reported jailbreak wasn’t unique to Fable 5. Multiple less capable models, including Opus 4.8 and GPT-5.5, identified the same vulnerabilities, and every model tested could reproduce the single exploit demonstration in the Amazon report. Anthropic trained a new safety classifier that now blocks the specific technique in over 99% of attempts, validated by researchers at the Commerce Department’s Center for AI Standards and Innovation. The trade-off is more false positives on routine coding and debugging, so expect Fable 5 to hand off to Opus 4.8 a bit more often than before.
• The price of restoration is a formal government relationship. As part of the resolution, Anthropic agreed to give designated agencies early access to future frontier models before public release, share threat intelligence, coordinate on future launches, and report malicious use. It’s also standing up a 24/7 jailbreak monitoring team, opened a HackerOne bounty program for Fable 5, and is drafting a shared framework with Amazon, Microsoft, and Google for scoring how dangerous a given jailbreak actually is.
• Claude Sonnet 5 launched June 30, and its restraint is part of the story. Sonnet 5 is Anthropic’s new everyday workhorse, a substantial upgrade over Sonnet 4.6 in reasoning, tool use, coding, and knowledge work. Anthropic says it approaches Opus 4.8 performance on some tasks at higher effort settings, for meaningfully less money. It’s the new default on Free and Pro plans and available everywhere, including Claude Code, with introductory pricing of $2 per million input tokens and $10 per million output tokens through August 31, after which it moves to $3 and $15. Two details are worth pointing out. First, Anthropic deliberately did not train Sonnet 5 on cybersecurity tasks, and its performance on dangerous cyber evaluations is far below Opus 4.8 and Mythos 5. Shipping a deliberately cyber-weak workhorse the same day the export controls lifted looks like a strategy: The Claude lineup is now visibly stratified by cyber capability, which is exactly the structure that satisfies a government-negotiated release regime. Second, Sonnet 5 uses a new tokenizer that maps the same input to roughly 1x to 1.35x more tokens, and the introductory pricing is set to make the transition roughly cost neutral. Model your costs on the standard price, not the promo.
• Claude Science launched at the June 30 science event. Claude Science is an auditable AI workbench for researchers, not a new model. It runs on existing Claude models including Opus 4.8, shows scientific artifacts alongside the exact code and environment that produced them, and uses a reviewer agent that flags bad citations and figures that don’t match their code as the pipeline runs. It can run on a lab’s own infrastructure and is in beta on macOS and Linux for Pro, Max, Team, and Enterprise plans. Dropping this two weeks after hiring Nobel laureate John Jumper makes Anthropic’s science ambition pretty clear.
• Claude Tag puts Claude on the team in Slack. This one launched June 23 and got buried under the export control news, but it is notable. Claude Tag lets Claude join Slack channels as a team member. Anyone can tag @Claude to delegate a task, and it works asynchronously, builds context from its channels over time, and can schedule its own follow-ups over hours or days. One Claude per channel means everyone sees what it’s working on and can pick up where a colleague left off. Anthropic says 65% of its own product team’s code is now created by its internal version. The enterprise controls are what IT leaders should note: Admins scope tools, data, and memories separately per channel, set token spend limits at the organization and channel level, and get a log of everything @Claude did and who asked for it. It’s in beta for Enterprise and Team plans, runs on Opus 4.8, and replaces the existing Claude in Slack app.

OpenAI: GPT-5.6 arrives, but the government picked the first customers

• GPT-5.6 Sol, Terra, and Luna launched June 26 in a government-gated preview. OpenAI released its next model family to roughly 20 organizations whose participation was shared with the US government at the administration’s request. Sol is the flagship at $5 per million input tokens and $30 per million output tokens, with a max reasoning mode and an ultra mode that coordinates subagents. Terra is the balanced tier at half Sol’s price, and Luna is the fast, cheap option at $1 and $6. OpenAI says Sol is its strongest model yet in coding, biology, and cybersecurity, and claims it edges Mythos 5 on some coding benchmarks while using far fewer output tokens. General availability is promised in the coming weeks.
• Sam Altman didn’t hide his frustration. He called the staggered release “reasonable” but “not optimal,” and OpenAI said government-approved rollouts shouldn’t become the long-term default. The gating follows the June 2 executive order asking frontier labs to submit models for government review up to 30 days before release. That 30-day window for agencies to define the review process closed July 2, so the formal framework should be public soon.
• A governance detail buried in the system card. OpenAI classified all three GPT-5.6 models at its “High” risk level for cyber and biological capability, not just Sol. That means even the cheap Luna tier may carry governance obligations for organizations using it in security or life sciences workflows. Read the system card before assuming the budget tier is the low compliance tier.
• The IPO may slip to 2027. OpenAI is reportedly leaning toward delaying its public offering until 2027, with SpaceX’s post-debut stock slide reportedly a factor in the thinking. Anthropic’s S-1 is already filed, so the two leading labs may now hit the public markets a year apart.

Google: The flagship slips again, and the exits keep coming

• Gemini 3.5 Pro missed June. Sundar Pichai told the I/O crowd in May to give Google until next month, and June closed with no launch. The model stays in limited Vertex AI enterprise preview, and Google now points to July, citing quality refinements after early tester feedback. The promised specs are still compelling, including a 2-million-token context window and a Deep Think reasoning mode, but this is Google’s second major delivery miss of the year. Don’t build your roadmap around a July date that Google itself won’t commit to.
• More researchers walked. Following the Shazeer and Jumper exits we covered last week, reports indicate four more senior Gemini researchers announced moves to Anthropic during the past week. A missed flagship deadline and a talent outflow at the same time should concern anyone betting on Google’s frontier roadmap.
• The accidental upside. With Fable 5 fresh off a government ban and GPT-5.6 launching through a government-approved list, Gemini is currently the only frontier lineup without access restrictions. That’s partly because Google’s shipping models haven’t crossed the informal capability thresholds drawing government attention. Whether 3.5 Pro changes that when it lands is an open question Google hasn’t addressed publicly.

Microsoft: The in-house coding model reaches the enterprise

• MAI-Code-1-Flash is generally available for Copilot Business and Enterprise. GitHub made Microsoft’s in-house coding model GA for managed plans on June 26, with admin policy controls and usage-based billing at $0.75 per million input tokens and $4.50 per million output tokens. Microsoft positions it for fast, iterative agentic coding rather than complex multifile work, with vendor-run benchmarks showing it well ahead of comparable small models. The pitch is efficiency, with a small, cheap, fast model for high-volume agent tasks now that Copilot bills by the token. Four weeks after Build, the post-OpenAI model strategy is now in enterprise production.
• Copilot’s model menu keeps widening. GitHub also added Claude Sonnet 5 (covered in the Anthropic section) to Copilot as generally available across Pro, Business, and Enterprise plans, running under zero data retention. Between MAI models, OpenAI, and Anthropic options in one interface, Copilot is becoming a model marketplace with Microsoft controlling the routing and the billing.

Amazon: A quiet week while Fable 5 comes back to Bedrock

• The AWS Weekly Roundup was light. AWS shipped an agentic CX designer for Amazon Connect, EC2 AMI watermarks, and took a seat on MySQL’s new open governance steering committee. Useful, but not headline material.
• The item that matters is pending. Anthropic says Fable 5 access on AWS is being restored as quickly as possible. If your Fable workloads run through Bedrock, confirm availability before restarting pipelines.

Our Take

The 19-day Fable 5 shutdown ended the way these things tend to end, with a deal. Anthropic got its model back and the government got what it wanted all along, including early access to future frontier models, threat intelligence sharing, and a seat at the table before launches. OpenAI read the room and offered the government a preview list before being ordered to. Whatever the July 2 executive order framework says on paper, the practice is already set. Shipping a frontier model in the US now involves Washington, and the interesting question is no longer whether the government reviews these models but how predictable and fast that review becomes. A repeatable 30-day process the industry can plan around is manageable. A discretionary one that can pull a shipped product offline for three weeks is a risk that belongs in every AI vendor assessment.

The other thing this week made plain is that the restoration doesn’t undo the lesson. For 19 days, enterprises that built on the most capable public model in the world had no access and no timeline. It came back with new classifiers, more false positives, a different pricing window, and pending hyperscaler availability, all of which need retesting before it earns back a place in production. Portability stopped being a nice-to-have this month.

What IT leaders should be doing:

• Retest before you readopt Fable 5. The restored model has a new safety classifier with a higher false positive rate on routine coding, and blocked requests silently hand off to Opus 4.8. Validate your workflows against the new behavior during the included usage window before July 7, and confirm Bedrock, Google Cloud, or Foundry availability before restarting production pipelines there.
• Add government review to your model release assumptions. GPT-5.6’s general availability date is now partly a government decision, and Anthropic has committed to prerelease government access for future models. Assume frontier launches can slip or gate on review, and don’t commit delivery dates that depend on a model shipping on schedule.
• Read the GPT-5.6 system card before choosing a tier. All three models carry OpenAI’s High risk classification for cyber and bio capability, including the budget Luna tier. If you operate in security, life sciences, or other sensitive domains, the compliance footprint is the same across tiers even though the price isn’t.
• Hold the line on Gemini 3.5 Pro. July is guidance, not a commitment, and it’s Google’s second slip this year. Evaluate Gemini 3.5 Flash, which is real and shipping, and treat Pro as upside.
• Turn the jailbreak severity framework into a procurement question. Anthropic, Amazon, Microsoft, and Google are drafting a shared standard for rating jailbreak risk. Ask your vendors whether they’ll adopt it, because a common severity scale would finally give you a way to compare vendor security claims.

Want to Know More?

• Big 5 AI Vendor Roundup: Week of June 22, 2026
• Claude Fable 5 Arrives With Mythos-Class Power and Prices
• Claude Mythos Preview and Project Glasswing: What IT and Security Leaders Need to Know Now
• GPT-5.5-Cyber: The Next Claude Mythos?
• Establish Your Adaptive AI Governance Program: From Principles to Practice