The same enterprise workload costs $4,811 on Anthropic's Claude, $3,357 on GPT-5.5, $1,071 on DeepSeek V4, and $544 on Zhipu's GLM-5.2. That's a ninefold gap from top to bottom for work that, increasingly, produces indistinguishable output.
Hold that number in mind while reading xAI's launch copy.
Because the number xAI wants you to see is a different one: Grok 4.5, live since July 9, priced at $2 per million input tokens and $6 per million output. That undercuts Opus 4.7 at $5/$25 and GPT's flagship tiers running $5 in and up to $30 out. On a launch slide, it looks like a haymaker.
It isn't. It's a model arriving late to a fight that was largely settled nearly four weeks earlier by a model you can download for free.
Last week we mapped a multi-polar frontier race where several labs traded the top of the leaderboard. This week the axis has shifted entirely: the story is no longer who holds the crown but who is collapsing the floor beneath everyone standing on the podium.
Consider the timing. Zhipu AI released GLM-5.2 on June 13, weeks before Grok 4.5 shipped. On SWE-bench Pro — the benchmark most developers actually trust to predict real-world code-repair performance — GLM-5.2 scored 62.1 against GPT-5.5's 58.6.
An open-weight Chinese model beat a US frontier flagship on the coding metric that matters, at one-sixth the API cost.
The pricing tells the rest. GLM-5.2 lists at $1.40 per million input tokens and $4.40 output. And its full 744-billion-parameter weights reportedly sit on Hugging Face under an MIT license — no usage restrictions, no regional limits, no API required at all.
Self-host it and your marginal cost per token is electricity.
This is the context that reframes Grok 4.5 from disruptor to latecomer. xAI didn't start a price war. It walked into one where the aggressive competitor gives its model away and lets you run it on your own hardware.
xAI didn't start a price war. It walked into one where the aggressive competitor gives its model away and lets you run it on your own hardware.
Now weigh the two capability claims side by side, because the asymmetry is the whole argument.
GLM-5.2's parity is independently observable. Its 62.1 sits on a public leaderboard, its weights are downloadable, and third parties — Graphistry, Semgrep — have run their own evaluations. Anyone skeptical can verify it in an afternoon.
Grok 4.5's parity rests on softer ground. The headline coding number — 64.7% on SWE-bench Pro — comes from xAI's own benchmarks. The company also claims 80 tokens per second and roughly 4.2x fewer output tokens than Opus 4.8 on the same tasks, again by its own measurement.
The most enthusiastic external endorsement so far is a YouTuber. Wes Roth built a 3D sailing game and a text RPG with the model and declared it had "caught up to Opus 4.8." Elon Musk called it "Opus-class."
That is not nothing. But it is a demo reel and a founder's tweet, not a public leaderboard you can audit.
So the two models make the same essential pitch — frontier-class coding at a fraction of frontier cost — and only one of them lets you check the work without taking the vendor's word for it.
This is the pivot that matters for anyone doing procurement math. When capability parity becomes the assumption rather than the differentiator, the frontier premium has to justify itself on something other than "we're better."
And increasingly it can't, because the "better" isn't unique and the price gap is enormous.
Return to that ninefold spread. The same workload: $4,811 on Claude, $544 on GLM-5.2. Even against GPT-5.5 at $3,357, GLM-5.2 delivers a higher SWE-bench Pro score for roughly a sixth of the cost.
That is not a discount you rationalize away with vague appeals to reliability. It is a line item that reshapes an entire AI budget.
Coinbase made exactly that call. The company moved internal workloads to open-weight models from Zhipu and DeepSeek and cut nearly 50% of its AI spending.
That is one of the largest public crypto firms deciding that the frontier premium wasn't worth paying for the tasks it was paying it on. Not a hobbyist optimizing a side project — a public company rewriting its cost structure.
The individual anecdotes rhyme with the enterprise number. An operations manager in San Diego watched an hourlong Claude coding session that cost about $10 drop below 50 cents on DeepSeek. Lindy's founder Flo Crivello moved his company off Anthropic and saved millions, with a line that functions as the whole thesis in miniature: "You don't need God to write your email."
If you can get lower tiers of intelligence for a tenth of the price, he argued, "it would be foolish not to do it."
That sentence is what the frontier premium is now up against. Not a claim that the expensive models are worse — a claim that their superiority, where it still exists, is no longer worth a 9x markup on most of the work being run.
The market is voting with tokens. DeepSeek variants alone reportedly reached 17% of OpenRouter token share, up from under 1% earlier. Six of the top ten models on a major leaderboard now come from China. GLM-5.2 is reportedly the only open-weight model competing on the Arena agent leaderboard alongside closed frontier systems.
Even David Sacks, hardly a booster of Chinese capability, called GLM-5.2 "as good as the currently available models from OpenAI and Anthropic" and placed it roughly level with GPT-5.5. Weeks earlier he'd pegged the US lead at six to nine months. The revision is the point.
None of this means the floor stops dropping. GLM-5.2 runs its training and inference on a cluster of roughly 100,000 Huawei Ascend chips — no Nvidia, no AMD, no Intel — a cost fact worth noting even as its supply-chain implications belong to another conversation. The relevant takeaway here is narrower: the low-cost tier is not a fragile arrangement dependent on subsidized access to Western silicon.
And the open-weight field is crowding fast. Tencent has open-sourced its 295-billion-parameter Hy3 MoE under Apache 2.0. Mistral has confirmed an open-weight flagship entering early access in July, explicitly targeting enterprises "currently defaulting to closed APIs."
Each new entrant applies the same downward pressure on the same premium.
The caveats are real and worth stating plainly. Open weights strip out safety guardrails as easily as they strip out cost, and regulated sectors — defense, finance, critical infrastructure — face genuine data-governance questions before routing anything sensitive through a Chinese API. GLM-5.2 also generates more output tokens per task than some peers, quietly eroding part of the sticker-price advantage in practice.
But those are frictions on adoption, not defenses of the premium. Self-hosting the weights sidesteps the data-sovereignty problem entirely, and even a token-inflated GLM-5.2 workload lands nowhere near Claude's bill.
Which returns us to the question our June 30 thread left open: do enterprises consolidate onto a single vendor, or build multi-vendor routing under cost pressure? The developer in Dallas already answered for himself, routing 90% of his tasks to DeepSeek, MiniMax and similar options and reserving Claude and ChatGPT for the remaining 10%.
That hybrid — cheap open weights for the bulk, frontier APIs for the sliver that genuinely needs them — is what the collapsing floor produces by default.
So Grok 4.5's $2/$6 is a fine price. It is simply no longer a strategy, because price is now table stakes and parity is now assumed.
The real casualty isn't GPT or Opus on any single benchmark. It's the premium itself — the idea that being marginally ahead on the hardest tasks entitles you to charge nine times as much for the ordinary ones. Cost and open weights, not benchmark supremacy, now set the developer default.
xAI arrived to that war with a competitive price and its own benchmarks. It arrived, all the same, after the terms had already been set by a model that costs nothing to download.