Three Frontier Models, One That Respects Your Data

2026-07-10

Published July 10, 2026. All pricing, benchmark scores, and data retention policies verified against official documentation and API behavior as of July 10, 2026.

In the span of eight days, three of the world's most capable AI models shipped. Grok 4.5 on July 8. GPT-5.6 Sol on July 9. Claude Fable 5 returned from government-enforced exile on July 1. Three frontier models. Three different answers to the same question: what do you owe the customer beyond raw capability?

The benchmarks are close enough that the choice between them shouldn't matter for most tasks. Depending on which benchmark you prefer, any of the three can claim a lead. TerminalBench 2.1 favors Sol (88.8%, 91.9% at Ultra). SWE-Bench Pro favors Fable 5 (80.3%). Both are within a few points of each other on most measures. Grok 4.5 hasn't published comparable scores but positions itself as a frontier coding and agentic model.

If this were a pure benchmark competition, you'd buy the cheapest one and move on. But the benchmarks aren't the story. The story is what each model costs, who can use it, and whether the company that built it trusts you with your own data.

Here's the scoreboard.

The Numbers

The price gap is stark. Sol costs half of Fable 5 on input and 40% less on output. Grok 4.5 is the cheapest frontier model available — 80% less than Fable 5 on input, 88% less on output. But price isn't the differentiator here. The data retention column is.

The Line That Matters

GPT-5.6 Sol is Zero Data Retention compatible. Let me say that directly because it's the difference between a model your IT department approves and one they block.

OpenAI's official announcement: "Programmatic Tool Calling lets GPT-5.6 write and run programs in-memory that coordinate tools and process intermediate results, making it Zero Data Retention (ZDR) compatible." The V8 sandbox that hosts tool execution is "architecturally isolated from OpenAI's data infrastructure." Enterprise ZDR agreements cover it.

This is not a footnote. It's the structural difference between a model that enterprises can deploy and one they can't. When Microsoft's own employees are blocked from using Fable 5 because of 30-day mandatory data retention — while the same Microsoft is deploying GPT-5.6 through Azure — the market has already decided the winner on terms. Not on benchmarks. On trust architecture.

Claude Fable 5 requires 30-day data retention for all Mythos-class models. No exceptions. Enterprise ZDR agreements, negotiated and paid for, are explicitly nullified. Anthropic's support documentation states this clearly: existing ZDR contracts "are not available under Zero Data Retention" for Fable 5. The model your legal team approved under a ZDR contract is one you can't use with that contract. It's not a pricing problem. It's a contractual impossibility.

Grok 4.5 has a stronger enterprise posture than its consumer-facing reputation suggests, but it's still less definitive than Sol's. xAI's Enterprise Terms of Service state that "All User Content will be automatically deleted," and Business plans ($30/user/month) include SOC 2 compliance and an explicit guarantee that data "will not be used for model training." However, the enterprise TOS contains a safety/AUP carve-out: xAI retains "the minimum data necessary" for policy enforcement and "will permanently delete it promptly when the justification no longer applies." This is a default-delete-with-exceptions model — not the unconditional ZDR that OpenAI offers on Sol or that Fireworks.ai offers on all models. It's worlds better than Fable 5's mandatory 30-day retention with no exceptions. But it's not a clean ZDR guarantee that an enterprise legal team can read once and approve.

The pricing story is dramatic. At $2/$6 with cached input at $0.50, Grok 4.5 is the cheapest frontier model by a wide margin — 80% less than Fable 5 on input, 88% less on output, and roughly 17x cheaper per task than Opus 4.8 according to one analysis. But the sleeper stat is max output: 500K tokens, matching its full context window. That's 4x more than Sol or Fable 5 at 128K. For long-form generation, document processing, or agent tasks that produce substantial intermediate output, Grok 4.5 gives you runway the other two don't.

The Irony

The irony is that Anthropic — the company that markets itself as the safety lab, the responsible one, the one that paused deployment for government review — has created the least enterprise-usable frontier model. The safety architecture that justifies Fable 5's 30-day retention is the same architecture that makes it impossible for enterprises to adopt. Anthropic built the safest model and made it unsafe for the customers who pay the most.

OpenAI — the company that spent years being criticized for moving fast, for shipping GPT-5 despite safety concerns, for being the "irresponsible" lab — shipped a frontier model that respects enterprise ZDR agreements on day one. The model everyone worried about is the one enterprises can actually use.

This is not about who has better safety values. It's about whose safety architecture aligns with enterprise procurement requirements. Anthropic's safety architecture says: we need to log your prompts for 30 days to monitor our classifier. OpenAI's safety architecture says: the sandbox is isolated, no data retained, enterprise ZDR covers it. The enterprise legal team reads both and approves one.

The Hubris

There's a word for what happens when a company becomes so dominant that it thinks it can dictate terms to the customers who built it. Anthropic reached that point sometime between the $965 billion valuation and the nullification of its own enterprise ZDR contracts.

Consider the sequence. Anthropic negotiates Zero Data Retention agreements with enterprise customers — contracts that cost real money, that legal teams reviewed, that procurement officers signed off on. These agreements say: your data is not retained. Your prompts are not logged. Your proprietary code stays yours.

Then Fable 5 ships. And Anthropic announces that those ZDR contracts — the ones it sold, the ones enterprises paid for — don't apply to this model. The word "nullified" is mine, but the substance is Anthropic's. The support documentation states that ZDR "is not available under Zero Data Retention" for Mythos-class models. Read that sentence again. It's a company telling its paying customers that the agreement they signed doesn't cover the product they want to buy, and no alternative exists.

This is not a technical limitation. OpenAI built a V8 sandbox that isolates tool execution from data infrastructure, making Programmatic Tool Calling ZDR-compatible on day one. Anthropic chose not to build equivalent isolation for Fable 5. The choice says: we're not going to engineer around this because we don't need to. You'll accept our terms because you don't have an alternative.

That was true in June. It's not true in July.

The hubris is visible in three decisions that would be reckless for any company and are self-destructive for one that markets itself as the responsible lab:

Nullifying signed contracts. Enterprise procurement doesn't forget this. When a vendor tells you the agreement you negotiated doesn't apply to their new product, procurement officers don't just reject the product — they start auditing the entire vendor relationship. Every contract. Every line item. Every assumption that Anthropic will act in good faith. The ZDR nullification didn't just block Fable 5. It burned trust across every Anthropic product those enterprises use.

The one-week free sample. Fable 5 is included in subscriptions at 50% weekly usage through July 7, then switches to metered credits at $10/$50 per million tokens. This is a pricing model designed to create dependency — give them the best model they've ever used for a week, then put it behind a meter. The chat participants called it the "drug addict" pricing model. Enterprise procurement officers have a different term: vendor lock-in. They've seen this playbook before, and they're not interested in repeating it.

No exceptions, no negotiation. Anthropic could have offered enterprise ZDR with a modified safety classifier. It could have created an enterprise tier with negotiated retention terms. It could have grandfatherd existing ZDR customers. It did none of these things. The posture is: take our terms or use a different model. For months, that posture worked because there wasn't a different model at the frontier.

There is now. Two of them.

The timing is what makes this hubris — not just arrogance, but arrogance at exactly the wrong moment. Anthropic spent 2025 and early 2026 building an unassailable market position. Opus 4.5 was the standard. Claude Code was the default coding agent. Fable 5 was the most anticipated model release in the industry. The company filed for IPO at $965 billion. Enterprise customers had nowhere else to go for frontier capability.

Then Sol shipped with ZDR at half the price. Grok 4.5 shipped with auto-delete at a fifth the price. And the enterprise customers Anthropic assumed were captive suddenly had exit ramps. The procurement officers who were told "take our terms" now have a Sol-shaped alternative. The legal teams who were told "your ZDR contract doesn't apply here" now have OpenAI's ZDR documentation to wave at their CFOs.

This is not about whether Fable 5 is a better model. It might be. The SWE-Bench Pro lead is real. The game-generation demos are extraordinary. But hubris is the belief that being the best excuses treating customers as captive. And in enterprise software, the best product loses to the buyable one every time.

The irony compounds: Anthropic's safety-first branding — the export control compliance, the jailbreak classifier, the 30-day retention for monitoring — is the architecture that made Fable 5 unbuyable. The company that built its reputation on responsibility shipped the least responsible enterprise product on the market. Not irresponsible in capability. Irresponsible in contract. And the customers are leaving.

The Contract Trap

There's a worse version of the hubris story, and it's the one enterprise procurement officers are living right now.

Through 2025 and early 2026, Anthropic's enterprise sales team was the most effective in the industry. They sold committed-spend agreements — multi-million-dollar contracts where enterprises guarantee a minimum annual token volume in exchange for discounted rates. The pitch was straightforward: lock in pricing now, get priority access to our roadmap, and rest assured you're partnered with the market leader. The subtext was even simpler: we're going to stay in front, and you're going to need what we're building.

The agreements are structured around committed annual spend negotiated up front, as documented by Morph's enterprise pricing analysis. You sign. You commit. You spend. The contract doesn't specify which models you'll use — it just says you'll spend the money on Anthropic products.

Enterprises signed these agreements when Anthropic was the unquestioned leader. Opus 4.5 was the standard. Claude Code was the default. Mythos was the promise on the horizon — the model that would extend the lead. The procurement calculus was conservative: bet on the market leader, lock in the rate, don't overthink it.

Then three things happened in sequence:

First, the export controls. On June 12, the U.S. government ordered Fable 5 and Mythos 5 shut down globally. Every enterprise customer lost access — not just to the new models, but to any model in the Mythos class. Gartner called it a "sovereignty warning bell" for CXOs. FifthRow documented enterprises "invoking poorly-defined 'event beyond our control' provisions, improvising crisis negotiations, and suffering prolonged legal and technical limbo." The models came back on July 1, but the contracts didn't account for an 18-day outage of the product you promised. Most force majeure clauses weren't written for "the government can turn off your AI."

Second, the ZDR nullification. When Fable 5 returned, enterprises discovered the model they'd been promised — the one their committed-spend agreements were supposed to cover — comes with mandatory 30-day data retention that nullifies their existing ZDR contracts. This isn't a pricing problem. It's a contractual impossibility. The enterprise has a legal obligation to their own customers (via SOC 2, HIPAA, client NDAs, or internal data governance policies) that prohibits 30-day retention of proprietary data. Anthropic is now telling them the model they committed to spend money on requires violating those obligations.

Third, the competition caught up. While Anthropic's enterprise customers were trapped between their minimum-spend commitments and their data retention requirements, OpenAI shipped GPT-5.6 Sol with ZDR at half the price. xAI shipped Grok 4.5 with auto-delete at a fifth of the price. The alternative Anthropic told enterprises they didn't have? It arrived.

Now the enterprise procurement officer faces an impossible position. Their committed-spend agreement with Anthropic is still on the books. They owe millions in token volume this year. But they cannot legally use Fable 5 — the model the agreement was pitched around — without breaking ZDR obligations. They can use Opus 4.8, which is fine but no longer competitive with Sol or Grok 4.5. They can use Sonnet 5, which burns through tokens so inefficiently that it costs more per task than Opus 4.8. They're legally committed to spending money on a vendor that no longer sells a product they can legally use at the frontier.

This is the trap Anthropic's sales machine built. Not intentionally — no one planned for export controls, and no one expected OpenAI to ship ZDR on day one. But intention doesn't matter when the contract is signed and the model is unusable. The enterprise response is already visible in the data: companies actively cutting Anthropic spend, Uber burning through AI budgets and reconsidering, Microsoft blocking Fable 5 internally. These aren't preference shifts. They're contractual escape attempts.

The committed-spend agreements were supposed to lock enterprises into Anthropic. Instead, they locked Anthropic into a product strategy that made its flagship model unbuyable — and gave its customers a very expensive reason to find the exit.

The Real Choice

If you're an enterprise evaluating these three models, here's what procurement actually sees:

GPT-5.6 Sol: Frontier capability, half the price of Fable 5, enterprise ZDR-compatible. Three tiers available (Luna at $1/$6, Terra at $2.50/$15) so you can route simple tasks cheaply and reserve Sol for hard problems. Tool calling works in an isolated sandbox. Your legal team can read the ZDR documentation and sign off.

Claude Fable 5: Frontier capability, excellent SWE-Bench Pro score, premium pricing. Requires 30-day retention of all prompts and outputs. Your existing ZDR contract is nullified. Microsoft blocked it internally. Your legal team won't approve it for anything involving proprietary code, customer data, or internal strategy — which is everything an enterprise AI model does.

Grok 4.5: Frontier capability at commodity pricing ($2/$6, $0.50 cached). SOC 2 on Business plans, enterprise TOS guarantees automatic deletion with a narrow AUP carve-out. The cheapest frontier model — roughly 17x cheaper per task than Opus 4.8 — with a 500K max output window that dwarfs the competition's 128K. Enterprise posture is real but "call us" — SSO, dedicated infra, custom compliance all require talking to sales.

The benchmarks don't decide this. The price doesn't decide this. The data retention column decides this. And in that column, one model has a checkmark, one has a strikethrough, and one has a question mark.

What Happens Next

The Fable 5 story has already entered enterprise procurement history as a case study in how to make an unbeatable product unbuyable. The model generates working video games from single sentences. It leads SWE-Bench Pro. It's revolutionary. And the largest enterprise software company in the world told its employees not to touch it within 24 hours of launch.

Anthropic's response to this will define the next phase of the market. If Fable 5 gets ZDR — if the safety classifier is refined enough that Anthropic can offer zero retention — it reclaims the enterprise. The model is good enough that customers will pay $10/$50 for it if they don't have to also surrender their data. If Anthropic holds the line on 30-day retention, Sol takes the enterprise market by default. Not because it's better. Because it's buyable.

Grok 4.5 is the wildcard — already priced to disrupt both Sol and Fable 5. At $2/$6 with $0.50 cached input and SOC 2 on Business plans, it's the cheapest frontier model that also clears a basic enterprise compliance bar. The routing architecture that practitioners are already building — cheap OSS for 70-80% of volume, frontier for the rest — gets a natural default. Grok for everything, Sol for the hard stuff with clean ZDR, Fable 5 for... the customers who haven't read their data retention agreements yet.

The three-way race was supposed to be about capability. It turned out to be about contracts. The model that wins isn't the one with the highest benchmark score. It's the one your legal team lets you use.

Benchmark data: GPT-5.6 Sol TerminalBench 91.9% Ultra, Claude Fable 5 SWE-Bench Pro 80.3%, head-to-head comparison. Pricing: GPT-5.6, Fable 5, Grok 4.5. ZDR: OpenAI GPT-5.6 Programmatic Tool Calling ZDR-compatible, Fable 5 30-day retention. Grok enterprise: Enterprise TOS (auto-delete), Business plan SOC 2 + no-training guarantee, Grok 4.5 $2/$6 cached $0.50, 17x cheaper vs Opus 4.8 analysis. Simon Willison on Sol pricing context: simonwillison.net.