Grok 4.5 Scores 64.7% on SWE Bench Pro, Undercuts Rivals at $2/1M Tokens

For AI developers, Grok 4.5 presents a compelling cost-performance trade‑off, offering decent code-fixing capabilities at a fraction of the cost of Anthropic's Opus 4.8 ($2 vs $5 per million input tokens). However, benchmark gaps of 5–6 percentage points on DeepSWE and SWE Bench Pro highlight that cutting‑edge reasoning still commands a premium.

On Wednesday, SpaceXAI released Grok 4.5, its first publicly available model since the SpaceX-xAI merger closed in February and as SpaceX progresses toward a $60 billion acquisition of developer tooling company Cursor. The model is aimed squarely at coders, engineers, and what the company calls 'knowledge workers' — a broad category spanning software developers, lawyers reviewing contracts, and finance teams building Excel models. But the headline isn't raw capability. SpaceXAI is pitching Grok 4.5 on price, speed, and internal validation at Tesla and SpaceX, with Elon Musk himself acknowledging that the model is at least one generation behind Anthropic's latest flagship.

“

That undercuts Anthropic's current primary flagship, Claude Opus 4.8, which runs $5 input and $25 output, and substantially beats OpenAI's just-launched GPT-5.6 Sol at $5 input and $30 output.

The pricing is the most aggressive differentiator. Grok 4.5 costs $2 per million input tokens and $6 per million output tokens. That undercuts Anthropic's current primary flagship, Claude Opus 4.8, which runs $5 input and $25 output, and substantially beats OpenAI's just-launched GPT-5.6 Sol at $5 input and $30 output. For cost-conscious enterprise developers building internal tools or iterating rapidly, that 60% to 75% discount on input costs could redirect volume away from premium-priced leaders. It's a classic price disruption play, reminiscent of how low-cost cloud providers grabbed market share from hyperscalers, but applied to frontier AI models where performance is usually the sole differentiator.

Benchmark results tell a story of strategic trade-offs. On DeepSWE 1.1, a benchmark that measures how reliably a model can close real-world software bugs submitted by developers, Grok 4.5 scored 53%, behind Opus 4.8 at 59% and GPT-5.5 at 67%, while Anthropic’s broader top-of‑the‑line Claude Fable 5 reached 70%. On SWE Bench Pro — a tough collection of software engineering problems — Grok 4.5 posted 64.7%, enough to surpass GPT-5.5 (58.6%) on that specific measure but still trailing Opus 4.8 (69.2%) and the towering 80.4% from Fable 5. These results confirm that Grok 4.5 is competitive with the previous generation of premium models, but cannot match current-gen reasoning depth.

Musk explicitly framed this gap on X, writing that Grok 4.5 is “roughly comparable to Opus 4.7, but much faster.” Opus 4.7 is the predecessor to Opus 4.8; by drawing that comparison, Musk set expectations for a model that is solid but not class-leading. The deliberate positioning echoes SpaceXAI's broader strategy of building vertically integrated AI that serves its own engineering‑heavy companies first and the market second. Engineers at Tesla and SpaceX have reportedly been using Grok 4.5 for months, providing a real‑world proving ground that the company now uses to sell to external developers.

The competitive landscape is heating up rapidly. OpenAI’s launch of GPT-5.6 Sol on the very same day, mere hours after Grok 4.5, shows that the premium coding‑model segment is now a battleground. Anthropic’s Fable 5 series still holds the performance crown, but its pricing likely remains at a premium. SpaceXAI’s pending acquisition of Cursor — a popular AI‑enabled code editor — hints at an ecosystem play: integrate a cheap, fast model directly into the tools developers already use, potentially locking in a developer base through deep workflow integration rather than sheer benchmark dominance.