Grok 4.5 Makes Token Efficiency the New Benchmark

An Opus-class claim at a warehouse-club price

Elon Musk did not launch a smarter model this week. He launched a cheaper argument.

On July 8, SpaceXAI released Grok 4.5, a model Musk described as “Opus-class” — internal evaluations, he said, put it roughly comparable to Anthropic’s Opus 4.7, but much faster. The claim matters less than the invoice attached to it. Grok 4.5 costs $2 per million input tokens and $6 per million output tokens, with cached input at $0.50 per million, per the model’s OpenRouter listing. Anthropic’s Fable 5 lists at $10 and $50. OpenAI’s GPT-5.5 sits at $5 and $30. SpaceXAI just parked a self-declared frontier model in the pricing lane of a mid-tier workhorse, and it did so weeks after going public.

The release is unusual in a second way: it was built with a customer. Cursor, the AI coding company, says Grok 4.5 was trained jointly with SpaceXAI on trillions of tokens of Cursor data — real interactions between developers, codebases, and software tools — then broadened with STEM tasks, research papers, and other knowledge work. Cursor calls it “the first we’ve built for more than software engineering.” That provenance is the story beneath the story. The most valuable training data for agentic models is no longer scraped text; it is telemetry from people doing real work inside real tools, and the companies that own that telemetry are becoming co-manufacturers of frontier models rather than mere distributors.

The stakes are easiest to see from the buyer’s chair. Two days ago I argued that cheap Chinese models were winning American workloads by clearing the quality floor and then competing on the invoice. Grok 4.5 imports that exact playbook into the American frontier tier. The Decoder’s analysis puts it bluntly: for a representative coding-agent task, Grok 4.5 costs $2.49, against $5.07 for GPT-5.5 in Codex and $11.80 for Fable 5. If those ratios hold across workloads, benchmark deficits of five to fifteen points stop being disqualifying and start being negotiable.

What is at risk, then, is not any single lab’s leaderboard position. It is the premium pricing structure of the entire frontier tier. Anthropic and OpenAI have spent two years training enterprises to believe that the best model is worth five to ten times the price of the second-best. Grok 4.5 is a direct attack on that belief, mounted by a company with its own 500k-token context window, 80-plus tokens per second of serving speed, and — crucially — its own power plants, launch infrastructure, and freshly raised public-market capital to subsidize the assault.

There is a quieter stake too: what counts as “frontier” is being redefined mid-race. On the independent Artificial Analysis Intelligence Index, Grok 4.5 scores 54 and ranks fourth, behind Claude Fable 5, GPT-5.5, and Claude Opus 4.8 — but it tops the field on agentic tool use, and the median model in its price tier scores 30. Fourth place at fifth-place prices used to be a consolation bracket. In an agent economy where models are hired by the task rather than admired by the benchmark, fourth place at those prices may be the most commercially important position on the board.

Tokens per task is the metric that pays the bills

The sticker price is the headline. The token count is the weapon.

SpaceXAI’s launch materials lean hard on one number: Grok 4.5 resolves agentic tasks with an average of 15,954 output tokens, versus 67,020 for Opus 4.8 in max mode — a 4.2x efficiency gap, per the company’s own figures summarized by MarkTechPost. The model reportedly solves tasks in under half the steps of comparable frontier models. Treat vendor efficiency claims with appropriate suspicion, but notice what kind of claim it is. This is not “we are smarter.” It is “we waste less of your money while being smart enough,” and that claim is checkable by any customer with a billing dashboard.

Stitch the efficiency figure to the price sheet and the real gap emerges. At $6 per million output tokens, Grok 4.5’s average resolved task carries roughly $0.10 of output spend. Opus-tier output pricing of $25 per million, applied to 67,020 tokens, yields about $1.68 per resolved task. That is a 17x per-task cost gap — four times wider than the 4.2x token figure alone suggests, because price and verbosity compound. This is the proprietary takeaway operators should carry into their next procurement meeting: multiply the price delta by the verbosity delta before comparing models, because vendors will only quote you whichever single number flatters them.

The benchmark picture, read honestly, supports “competitive” rather than “leading.” On Terminal Bench 2.1, the three frontier contenders cluster within a single point — Fable 5 at 84.3%, GPT-5.5 at 83.4%, Grok 4.5 at 83.3%. On SWE Bench Pro, Grok 4.5’s 64.7% beats GPT-5.5’s 58.6% but trails Fable 5’s 80.4% by a wide margin. On DeepSWE 1.1, Grok manages 53% against Fable’s 70%. The pattern is consistent: wherever tasks are long, tool-heavy, and terminal-shaped — the Cursor training distribution — Grok 4.5 plays at the top table. Wherever tasks demand deep repository comprehension, Anthropic’s lead survives intact.

Speed compounds the economics. Artificial Analysis measures Grok 4.5 at 89.5 output tokens per second, above the 76.5 t/s median for reasoning models in its price tier — and because it emits fewer tokens per task, effective wall-clock latency per resolved task drops even further than the serving speed implies. For interactive agent products, where a user is watching a spinner, tokens-per-task and tokens-per-second multiply into the metric that actually governs experience: seconds per answer. A model that thinks in shorthand beats a model that thinks in essays, even at equal intelligence.

The Cursor angle deserves its own weight. Cursor’s Composer models proved that a code-editor company could train competitive coding models; Grok 4.5 extends the thesis to knowledge work generally, with reinforcement learning run against realistic environments constructed by a distributed agent system. This is the same structural insight OpenAI reached from the opposite direction with GPT-5.6’s subagent orchestration, which I covered in yesterday’s piece on the agent race: the frontier is no longer a model, it is a training loop wrapped around real work. xAI needed distribution and workflow data; Cursor needed frontier-scale compute and a base model. The merger of those needs produced a model neither could have built alone.

Distribution is the final piece of the evidence, and it launched simultaneously with the weights. Grok 4.5 arrived day-one across Cursor’s desktop, web, iOS, CLI, and SDK surfaces, with individual and team plans carrying doubled usage for the first week and a “fast” variant priced at $4 per million input and $18 per million output for latency-sensitive work. Compare that to the traditional frontier launch — API first, integrations trickling out over a quarter — and the strategy is legible. SpaceXAI is not waiting for developers to come evaluate the model; it is inserting the model into the editor where a large share of the world’s agentic coding spend already lives, priced to make switching the default rather than the experiment.

And the market context makes the timing pointed. The industry has spent 2026 relearning thrift — I traced that shift in The AI Token Binge Is Over, where the argument was that enterprises had begun rationing tokens the way cloud teams once rationed compute. Grok 4.5 is the first frontier-tier release engineered, priced, and marketed for exactly that buyer. It does not ask customers to believe in artificial general intelligence. It asks them to believe in their own cost dashboards, which is a much easier sell in a quarter when CFOs are auditing AI spend line by line.

Cheap, fast, and confidently wrong

Every discount has a denominator. Grok 4.5’s is trust.

Start with the number SpaceXAI did not put in the launch post. Independent testing summarized by Tech Times found Grok 4.5’s hallucination rate more than doubled from its predecessor, rising from 25% to 54%, alongside 0.63 guardrail violations per task — both worst in their comparison sets. The pattern is a known one: larger, more capable models know more and assert it with more unwarranted confidence. For a code-review agent whose output gets executed and tested, a hallucination is an inconvenience the harness catches. For legal, financial, or client-facing work — precisely the “knowledge work” market this model was built to enter — a model that is wrong 54% of the time when it ventures beyond its knowledge, and sounds certain while doing it, is a liability with an API key.

The benchmark provenance has its own asterisks. Cursor disclosed that Grok 4.5 enjoyed an unfair advantage on its own CursorBench because earlier snapshots of test codebases accidentally leaked into training data, a contamination it says it later removed. Credit the honesty; note the implication. When your training corpus is the same telemetry your benchmarks are built from, contamination is not an accident waiting to happen — it is the default state requiring active prevention. Buyers should also notice that xAI’s launch materials cite DeepSWE 1.0, where Grok beats Opus 4.8, while independent replication on DeepSWE 1.1 shows Grok trailing Fable 5 by seventeen points. Version selection is the new benchmark shopping.

Then there is the bias file, which no enterprise procurement review will skip. Promptfoo’s systematic evaluation of political bias in Grok models documented a system with a history of edited system prompts and outputs that drift toward its owner’s views — followed by apparent overcorrection that made the model unnecessarily critical of its own parent company. The specific direction of the tilt matters less than the demonstrated fact of the dial. A vendor that has visibly turned the knob on contested outputs, in both directions, within eighteen months, is asking regulated customers to price in the possibility that the knob turns again — mid-contract, without a changelog.

The economics invite skepticism too. SpaceXAI is a weeks-old public company burning capital across launch vehicles, satellite constellations, and gigawatt-scale data centers. A $2/$6 price on a frontier-adjacent model may reflect genuine architectural efficiency — the mixture-of-experts design and terse reasoning style are real — or it may reflect an introductory subsidy designed to buy market share before the next earnings call, the classic pattern of every venture-funded land grab from ride-sharing onward. Customers who rebuild their cost models around today’s Grok pricing should ask what the price looks like when the growth story needs margin. The cached-input discount and the pricier “fast” variant at $4 input and $18 output already sketch the shape of future segmentation.

Finally, the thesis itself — good-enough-plus-cheap beats best — has a failure mode: agentic work compounds errors. A model that resolves tasks at 83% on terminal benchmarks but hallucinates at the tail will complete nine steps of a ten-step workflow and poison the tenth. The cost of that failure is not $2.49; it is the engineer-hour spent discovering it, multiplied by every downstream system that consumed the bad output. Anthropic’s premium has always been an insurance premium. Whether it is worth 5x depends entirely on how expensive your failures are — which is why the frontier price war will be fought workload by workload, not model by model.

The floor is falling — position accordingly

Frontier intelligence just got a list price, and list prices only move one direction once a credible discounter shows up.

The near-term sequence is predictable because we have watched it twice this year. Chinese labs established that the quality floor could be cleared at a tenth of the price; incumbents responded with cheaper tiers and routing stories. Now a domestic, publicly traded, politically connected challenger has made the same move inside the frontier bracket, with none of the procurement friction that kept Chinese models out of regulated American workloads. Expect Anthropic to defend Fable 5’s premium with reliability data rather than discounts — its 80.4% on SWE Bench Pro is the number that justifies the invoice — and expect OpenAI to respond structurally, pushing GPT-5.6’s tiered family and router so that the price comparison becomes a portfolio comparison, as I argued yesterday.

The deeper shift is in what gets measured. For three years the industry’s scoreboard was a leaderboard; the last week suggests the next scoreboard is a receipt. Token efficiency — tokens consumed per task resolved — is emerging as the metric that unifies intelligence, verbosity, and price into the only number a business actually experiences. SpaceXAI’s decision to lead its launch with a 4.2x efficiency claim rather than a benchmark crown, and The Decoder’s conclusion that the benchmark gaps may simply not matter at these prices, both point the same direction. Labs will start training for terseness the way they once trained for test scores, because terseness is now revenue.

Watch the Cursor precedent hardest. If a tool company’s interaction telemetry can co-produce a frontier model, then every company sitting on dense workflow data — Figma, Notion, Databricks, Bloomberg, Epic — is a potential model co-manufacturer, and the labs’ scarcest input stops being GPUs and starts being partnerships. That inverts the last two years of power dynamics, in which application companies feared being steamrolled by the platforms beneath them. The application layer, it turns out, owns the one dataset the platforms cannot scrape: what expert users actually do, step by step, when the work is real.

For operators, the playbook writes itself, and it starts with refusing to treat any of this as settled:

• Re-benchmark on your own tasks within two weeks. Grok 4.5’s profile — elite on terminal-shaped agentic work, weaker on deep repo comprehension, worst-in-class on hallucination — means aggregate scores will mislead you in both directions. Run your top five workflows head-to-head and measure cost per resolved task, not cost per token.
• Add tokens-per-task to every model evaluation. The 17x per-task gap between Grok 4.5 and Opus-tier pricing came from multiplying two deltas most teams track separately. Your eval harness should log output tokens per completed task as a first-class metric alongside accuracy.
• Price the failure, not just the task. Route high-blast-radius work — legal drafting, financial analysis, anything client-facing — toward models with the best calibration data, and let the 54% hallucination number veto Grok there until independent testing says otherwise.
• Negotiate with the new floor. Even if you never deploy Grok 4.5, its $2/$6 price is now your leverage in every renewal conversation with Anthropic and OpenAI. Frontier vendors have been repricing quarterly; make sure your contracts can too.
• Audit your telemetry’s worth. If Cursor’s interaction data was valuable enough to co-train a frontier model, ask what your product’s workflow exhaust is worth — and read your existing AI vendor agreements to find out who owns it today.
• Assume the subsidy ends. Model any Grok-based cost structure at 2x current list price and confirm the economics still work. Introductory pricing from a company that needs a growth narrative is a promotion, not a promise.

Three signals will tell you whether this launch was a repricing event or a footnote. First, watch whether Anthropic or OpenAI touch list prices before September; a matched cut would confirm that the frontier premium is now contestable, while silence would suggest their enterprise books are stickier than the discourse implies. Second, watch the independent hallucination and calibration retests — if the 54% figure replicates at scale, Grok 4.5’s addressable market shrinks to workloads with cheap failure, whatever the price. Third, watch SpaceXAI’s first full earnings report as a public company, where the gap between Grok’s serving costs and its $2/$6 sticker will finally acquire a paper trail.

The frontier used to be defined by what only one lab could do. As of this week, it is increasingly defined by what several labs can do and what each charges for it. That is not the end of the intelligence race — Fable 5’s benchmark lead is real, and the next capability jump will reset the board again. But it is the beginning of the frontier’s commodity phase, and commodity phases reward a different discipline: not picking the smartest model, but knowing, per task, exactly how much intelligence you need and refusing to pay for more.

In other news

Meta opens its model API for business — Meta Superintelligence Labs released Muse Spark 1.1, a multimodal agentic model with a 1M-token context window, and opened the Meta Model API to developers in public preview at $1.25 per million input tokens and $4.25 per million output tokens, with $20 in free credits. The release puts Meta directly into the paid-API business model of OpenAI and Anthropic for the first time (TechCrunch).

Claude Cowork escapes the desktop — Anthropic expanded Claude Cowork to mobile and web, letting Max subscribers start agent tasks at a desk and monitor them from a phone even with the laptop closed. Anthropic’s sample of 1.2 million Cowork sessions across 600,000+ organizations found software development is just 8.7% of usage, with business-process work leading at 33.4% (TechCrunch).

Illinois writes the strictest state AI law yet — Governor JB Pritzker signed SB 315, the AI Safety Measures Act, making Illinois the first state to require annual independent third-party audits of frontier models from developers with over $500 million in annual revenue, plus 72-hour reporting of critical safety incidents. The law, which OpenAI and Anthropic both supported, takes effect January 1, 2028 (Capitol News Illinois).

Goldman bets $110M on AI loan officers — Taktile raised a $110 million Series C led by Growth Equity at Goldman Sachs Alternatives, with Tiger Global participating, to expand its agentic decision platform that automates underwriting, claims, and fraud decisions for institutions including Mercury and Monzo. The company reports 95% automation in B2B underwriting for some customers and has raised $184 million to date (Fortune).

Chinese giants fund AI glasses — Even Realities Technology raised $150 million in pre-Series B funding from Meituan and Tencent to scale its waveguide-optics AI smart glasses, one of the week’s largest hardware rounds. The deal signals that Chinese platform companies see wearables as the next distribution surface for AI assistants (Tech Startups).