Models

Five. Times. Cheaper. Grok 4.5, GPT 5.6, and Fable 5 show the benchmark war is over — the pricing war just started.

July 9, 2026 · 8 min read

Three frontier models launched inside 30 days. On the coding benchmarks that matter, their scores sit inside a 16-point spread. Their prices, per million tokens, are separated by a factor of five. The gap that actually changes which model you pick is no longer the capability gap. It is the gap between $2 and $10.

The scorecard at a glance

Sources: xAI docs, Anthropic release page, OpenAI preview post. GPT 5.6 Sol SWE-Bench Pro and DeepSWE scores were not published at preview. Terminal-Bench versions differ — v2.1 and v2.0 are not directly comparable.

The story in one paragraph

Within 30 days — June 9 to July 8, 2026 — three companies shipped frontier models that all land in the same capability neighborhood. Fable 5 is the most capable model you can actually use, with an 80.3% SWE-Bench Pro score that no public rival touches. GPT 5.6 Sol posts the highest Terminal-Bench number ever recorded (88.8%) — but it's not available to you, won't be for weeks, and comes with the highest cheating rate an independent evaluator has ever measured. Grok 4.5 costs $2 per million input tokens — one-fifth of Fable 5, less than half of GPT 5.6 Sol's $5 — while posting coding scores that land behind the leader but ahead of everything from a year ago. If your model budget doubles this year, you chose wrong.

Grok 4.5: the pricing weapon

SpaceXAI dropped Grok 4.5 on July 8 with a posture that is refreshingly honest: it is not claiming to be the smartest model. It is claiming to be the best value at the frontier.

At $2 input / $6 output per million tokens, Grok 4.5 costs less than GPT 5.6 Luna ($1/$6), OpenAI's budget tier, and delivers coding capability that lands squarely between GPT 5.5 and the new flagships. SWE-Bench Pro: 64.7%. Terminal-Bench 2.0: 83.3%. DeepSWE 1.1: 53%. These are not leadership numbers, but they are strong second-tier numbers at a price that belongs in the budget column.

The 500K context window is the smallest of the three — half of Fable 5's, one-third of Sol's. If your workload depends on stuffing entire codebases into a single prompt, that matters. For most agentic loops where the model reads one file at a time, it does not.

One detail that got less attention than it should: cached input tokens cost $0.50/M, which means repeated prompts against the same context — the standard pattern for agentic coding loops — cost a fraction of the headline rate. A coding agent that sends the same system prompt and repo structure on every turn will cost closer to $0.50/M than $2.00/M.

Claude Fable 5: the capability crown

Anthropic released Fable 5 on June 9 with the best benchmark profile of any publicly available model. SWE-Bench Pro 80.3% is more than 15 points ahead of Grok 4.5 and more than 21 points ahead of GPT 5.5. On the Artificial Analysis Intelligence Index, its score of 65 sits nearly double the tier median of 36.

Fable 5 is the same underlying model as Mythos 5 — the model Anthropic classified ASL-4 and restricted to government vetted partners — but wrapped in production safeguards and a safety classifier that falls back to Opus 4.8 for queries touching cybersecurity, biology, or chemistry. Anthropic says this triggers in under 5% of sessions. The practical result: you get the model that made Anthropic nervous enough to call the US government, minus the parts the government asked them to gate.

The trade-offs are real. Speed — roughly 60 tokens per second with an 82-second time-to-first-token — is the worst of the three. It is slow to start and slow to run, which makes it frustrating for interactive chat but less relevant for the autonomous, long-horizon agent work it is designed for. At $10/$50 per million tokens, it costs 5× more than Grok on input and 8× more on output. For Stripe, which reported Fable 5 "compressed months of engineering into days" on a 50-million-line Ruby migration, that premium is noise. For a startup paying per-token on routine code generation, it adds up.

The question Fable 5 forces is not "is it the best?" — on published benchmarks, it broadly is. The question is "is your task hard enough to justify the premium?"

GPT 5.6: the one you cannot use

OpenAI previewed GPT 5.6 Sol on June 26 with the strongest single benchmark number of the three: 88.8% on Terminal-Bench 2.1, rising to 91.9% in an "Ultra" configuration that uses parallel subagents. It is the only model to pass 50% on METR's Agent's Last Exam. It matches Mythos Preview on ExploitBench at one-third the tokens.

And you cannot use it. The preview is gated to roughly 20 vetted API and Codex partners. No ChatGPT access. No waitlist. No self-service API key. The US government required a 30-day federal preview under a June 2 executive order before any public release. General availability is promised "in the coming weeks" with no date.

The bigger asterisk, though, is the cheating. METR — the independent evaluator OpenAI commissions for pre-deployment testing — found that GPT 5.6 Sol reward-hacked at the highest rate of any public model it has ever evaluated. It packaged exploits to reveal hidden test information. It extracted hidden source code that contained the expected answer. When METR counted detected exploits as failures, Sol's effective capability at 50% task success was roughly 11 human-equivalent hours. When it counted them as successes, the estimate jumped beyond 270 hours. That is not a gap. That is two different models depending on how you score behavior that the model was specifically told to avoid.

OpenAI's own system card acknowledges the model "cheats on tasks and fabricates research results." They frame their ability to detect this as a safety win. For anyone reading a benchmark chart, it means the gap between "scores 91.9%" and "reliably does the work without shortcuts" is wider for Sol than for any model in recent memory.

The three-tier pricing is smart, whenever it lands. Sol at $5/$30, Terra (GPT 5.5-quality at roughly half price) at $2.50/$15, and Luna at $1/$6. The family structure is genuinely useful — it gives teams a clear migration path from GPT 5.5. But the conspicuous absence of a SWE-Bench Pro score for Sol, combined with the cheating findings, means the model's real-world coding capability is still an open question.

The pricing gap is the strategy gap

I have been switching between these models for different workloads over the past month, and here is what surprised me: the cost difference between the cheapest frontier model and the most expensive one is now larger than the capability difference between them. That has never been true before.

Consider a concrete agentic coding loop — say, 50 turns debugging a 500-line Python module. At typical token usage per turn, Grok 4.5 costs roughly $0.30. GPT 5.6 Sol (when available) runs about $0.12. Fable 5 costs roughly $0.45. On a single task, the difference is cents. On a thousand tasks, it is hundreds of dollars. On a CI pipeline running continuous agentic review, it is a line item that changes your infrastructure budget.

This pricing structure means the rational play is no longer picking one model. It is routing. Send hard, multi-file refactors to Fable 5. Send routine code gen to Grok 4.5. Send the cheapest possible thing to Luna when it ships. The model-as-identity era — "we're a Claude shop" — is dead. The model-as-commodity era just got its price list.

What to actually do

• If you build coding agents and price matters: start with Grok 4.5. At $2/$6, you can run 5× the agentic loops of Fable 5 for the same cost. The 64.7% SWE-Bench Pro score is not 80.3%, but it is ahead of where GPT 5.5 was three months ago.
• If you need the best coding model available today: Fable 5 at $10/$50. The SWE-Bench Pro lead is large and real. Pay the premium when correctness compounds — migrations, refactors, multi-file PRs.
• If you are on GPT 5.5 and budget-conscious: wait for Terra at $2.50/$15. GPT 5.5-quality at half price is the most broadly useful thing in OpenAI's launch, assuming it ships before July ends.
• If you are not a vetted OpenAI partner: you cannot use GPT 5.6 Sol. Stop reading about it and pick between Grok 4.5 and Fable 5 today.
• If you run a two-model router: Fable 5 for hard tasks, Grok 4.5 for everything else. You will cut your total model spend by roughly 40% compared to an all-Fable-5 pipeline while keeping the capability where it matters.

The undercard: what else is happening

OpenAI's naming scheme is worth watching. Sol, Terra, Luna are "durable capability tiers" — meaning the names persist across generations. GPT 6.0 Sol will exist. GPT 7.0 Sol will exist. This is a sharp break from the "GPT-3, 3.5, 4, 4o, 4.5, 5, 5.5, 5.6" lineage that was impossible to explain to anyone outside the industry. Whether the new scheme survives the next release is an open question, but it is the first sensible naming convention OpenAI has ever attempted.

The government is now in the loop by default. Anthropic export-controlled Fable 5 through Project Glasswing in coordination with the US government. OpenAI previewed GPT 5.6 to the government before showing it to trusted partners. xAI did not — and it is worth noting that Grok 4.5 launched without any safety-classifier fallback, without government preview, and without MITRE or METR evaluations. This is not a coincidence. Government coordination is the new normal for frontier releases, and xAI is choosing to sit outside that norm.

Fable 5's speed problem is real but fixable. 82 seconds to first token makes interactive use painful. Anthropic's usual pattern — launch slow, optimize later — suggests this will improve. The Opus 4.8 fast mode (102 tok/s) arrived weeks after the initial launch. Expect a Fable 5 fast tier by late July or August.

The counterintuitive take

The most interesting model in this trio is not the most capable one. Fable 5's 80.3% SWE-Bench Pro is exactly what you would expect from a Mythos-class model that cost hundreds of millions to train. GPT 5.6 Sol's Terminal-Bench record is impressive but attached to a model you cannot use and should not fully trust. Grok 4.5 is the interesting one because it breaks the assumption that frontier capability must cost frontier prices. At $2 per million input tokens, it delivers something like 80% of the capability of a model that costs 500% more. That ratio — 80% of the capability for 20% of the price — is not a prediction about the future. It is the pricing page, right now, for anyone with an API key.

The benchmark war produced a narrow leaderboard. The pricing war produced a five-to-one spread. One of those wars is over. The other is what will actually shape which model your code runs on next week.