Three frontier models landed in a 36-hour window this month, and for the first time the cheapest option for a coding agent isn't the weakest one. If you're deciding what backend runs your agent's next thousand hours of tool calls, the choice is no longer "cheap and dumb vs. expensive and smart" — it's genuinely contested on the merits.

The takeaway: Grok 4.5 is the new value leader because it's cheaper per token and burns fewer tokens per task — those two effects compound. Opus 4.8 still wins on raw capability and is the only one of the three fully available in the EU today. GPT-5.6 gives you a tier menu to buy down cost without switching providers, but its cheapest tier tops out at the same $6/1M output price Grok 4.5 charges across the board.

What each one is actually optimizing for#

Grok 4.5 (xAI, released July 8) is xAI's first model built specifically for coding and agentic work, trained on real Cursor session data rather than generic benchmarks. It's priced at $2.00/1M input, $0.50/1M cached input, and $6.00/1M output — over 60% cheaper than Opus 4.8 or GPT-5.5 on input, and roughly 76% cheaper on output than Opus 4.8. It ships a 500K-token context window and a reasoning_effort dial (low/medium/high, high by default) that lets you trade latency for depth per call. It's built on a reported 1.5-trillion-parameter foundation and speaks both the Responses API and Chat Completions, so it drops into an OpenAI-shaped codebase with a base-URL change.

GPT-5.6 (OpenAI, released July 9) isn't one model — it's three: Sol, Terra, and Luna: Sol (flagship, $5/$30, matches GPT-5.5's old pricing), Terra ($2.50/$15), and Luna, a new $1/$6 production tier aimed at high-volume workloads. All three share a 1M-token context window and 128K max output. Altman pitched the family on a 54% token-efficiency gain over GPT-5.5, plus a 90%-off cache-read discount across tiers. On SWE-Bench Pro, Sol scores 64.6%, Terra 63.4%, Luna 62.7% — all a step up from GPT-5.5's 59.4%.

Opus 4.8 (Anthropic, released May 28) is the incumbent flagship: $5/1M input, $25/1M output at standard speed — unchanged from Opus 4.7 — with a 1M-token context window by default and a 3x-cheaper Fast Mode ($10/$50) for latency-sensitive work. It still leads the pack on SWE-Bench Pro at 69.2% and on Artificial Analysis's broader Intelligence Index (56 vs. Grok 4.5's 54).

The worked example: cost per completed task, not per token#

Sticker price alone understates the gap. On SWE-Bench Pro, xAI reports Grok 4.5 uses an average of 15,954 output tokens per task, versus roughly 67,020 for Opus 4.8 — about 4.2x fewer tokens to get to a finished diff. Independent testing from Artificial Analysis corroborates the direction, though with a slightly smaller gap.

Do the multiplication:

That's roughly a 17x difference in cost-per-completed-task — the price gap (4.2x) and the token-efficiency gap (4.2x) stack multiplicatively rather than canceling out. (We went deeper on that token-efficiency gap when Grok 4.5 first shipped.) This is the actual insight behind July's reset: a lower per-token price is a discount, but a lower per-token price combined with fewer tokens per task is a different cost curve entirely. That combination is what makes Grok 4.5's pitch different from "we're the budget option."

We don't have a public output-tokens-per-task figure for GPT-5.6 on SWE-Bench Pro, so the fairest comparison there is pass rate against list price: Sol matches Opus 4.8's price on input ($5) but nearly matches Grok 4.5's pass rate (64.6% vs. 64.7%) at five times the output price. Luna, the new $1/$6 tier, undercuts Sol's price steeply but gives up two points of pass rate versus Sol.

Context, tool-use, and where you can actually deploy#

Context window favors GPT-5.6 and Opus 4.8 (1M tokens) over Grok 4.5 (500K) — a real constraint if your agent needs to hold a very large repo or long transcript in-context rather than retrieving it. On agentic tool-use specifically, Artificial Analysis calls Grok 4.5 the single best agentic tool-use result on its board, and its Coding Agent Index score (76, run in Grok Build) lands on par with GPT-5.5 running at high effort in Codex — at a fraction of the token cost. Opus 4.8 still tops GDPval-AA v2's agentic Elo ranking outright.

Availability is the one constraint no price comparison fixes: Grok 4.5 is not yet live in the EU, either through xAI's own products or its API console, with EU availability targeted for mid-July 2026. GPT-5.6 and Opus 4.8 are both already available there.

Pick X if…#

Pick Grok 4.5 if your agent runs high volumes of tool-call-heavy tasks where cost-per-task compounds fast, you're outside the EU (or can wait a couple weeks), and you can tolerate a real but modest accuracy gap (64.7% vs. 69.2% on SWE-Bench Pro) in exchange for roughly 17x lower cost per completed task.

Pick GPT-5.6 if you're already deep in the OpenAI ecosystem and want the ability to dial cost up or down by tier — Luna for high-volume, low-stakes agent steps, Sol for the hard cases — without touching your provider integration.

Pick Opus 4.8 if you need the highest raw pass rate on genuinely hard, multi-file changes, you're building for EU users today, or the cost of a wrong agent output (a bad production diff, a broken migration) is high enough that a 4-to-5-point accuracy edge is worth 17x the token bill.