Let's set the division of labor first. Our GPT-6 Astra flagship review (/en/posts/gpt-6-astra-flagship-review) answers "which model is smarter." This article has a strictly narrower job: money only, no capability rankings. Who wins on reasoning quality or agent reliability is that article's problem. Here we answer one question: for the same workload, how much does the bill differ across today's flagship models? All prices and benchmark numbers below are official vendor figures as relayed through media and AI-tool aggregators; we have not re-run any of them, and we flag the provenance at every step.
This round of flagship price cuts is unusually dense with information. OpenAI shipped two price tiers at once with Sol and Luna. Anthropic cut both list prices and cache prices on Opus 5.5. xAI held Grok 4.7's prices flat but kept the lowest output rate in the dollar-denominated field. DeepSeek V4.1 Flash still bets on time-of-use pricing. Judging any single number in isolation will mislead you; you have to stack list price, cache discount, and cost-per-point into one picture.
1. Money per unit of capability: the honest way to compare
Raw list prices are the easiest numbers to misuse. A model with $2 input can end up cheaper in practice than a $1-input model that offers no cache discount, if its cache hit rate runs high. So our core metric here is not "dollars per million tokens" but "money per unit of capability" — how much workload completion each dollar buys.
Two ground rules before we start.
First, "capability" means different things in different tasks. Coding agents, cross-app workflow automation, and computer-use tasks probe different sides of a model, and the same model can flip positions across benchmarks. So we refuse to compress scores into one total. We place them, benchmark by benchmark, next to the prices.
Second, "money" is more than list price. For agentic workloads, the real bill equals input fees plus output fees, adjusted by your caching strategy and the repeated spend of failed retries. Cache pricing is the part that moved the most in this round of announcements, and the part most often ignored.
2. List prices: the full rate card
| Model | Input ($/M tokens) | Output ($/M tokens) | Cache read | vs. previous generation |
|---|---|---|---|---|
| GPT-6 Sol | 2 | 10 | 90% discount (approx. $0.20) | Price cut 50% |
| GPT-6 Luna | 0.10 | 0.50 | 90% discount (approx. $0.01) | 50% cheaper |
| Claude Opus 5.5 | 4 | 20 | Cut 60% ($0.50 down to $0.20) | List prices cut 20% |
| Grok 4.7 | 2 | 6 | No public discount data | Unchanged |
| DeepSeek V4.1 Flash | Off-peak 1 CNY / peak 2 CNY | Off-peak 4 CNY / peak 8 CNY | Not given in our source material | — |
Four structures jump out.
One: Opus 5.5's input and output rates are exactly 2x Sol's. That multiple is stated directly in the vendor announcements, not computed by us. There is also a Fast mode at $8/$40, officially described as 2.5x speed; we leave it aside here.
Two: Grok 4.7's $6 output rate is the lowest output price among the dollar-denominated models. For output-heavy tasks — long-form generation, large code dumps — that number carries a lot of weight.
Three: Luna's $0.10/$0.50 is not "budget flagship" territory anymore; it walks straight into DeepSeek's home turf. The relayed announcement draws the comparison explicitly: DeepSeek V4.1 Flash charges 1 CNY off-peak and 2 CNY at peak for input, 4 CNY and 8 CNY for output; Luna works out to roughly 0.7 CNY input and 3.5 CNY output (using $0.10 = approx. 0.7 CNY, a conversion rate as relayed by AI-tool aggregators; the two currencies are different pricing systems, so treat this as an order-of-magnitude reference, not a direct comparison). Luna is cheaper with no peak/off-peak swing — but DeepSeek's DeepSWE score of 74.2% beats Luna's 66.6%; a lower price is not a higher score.
Four: DeepSeek is the only vendor in this lineup still committed to time-of-use pricing. Off-peak rates are real money for batch users who can schedule work; for real-time chat users they effectively always pay peak.
3. The real bill after cache discounts
Cache pricing moved harder than list pricing in this round. Sol and Luna give a 90% discount on cached input reads; Opus 5.5's cache read price drops from $0.50 to $0.20 (a 60% cut); Grok 4.7 has no published cache discount. That deserves its own table.
Method: all figures below are worked examples from official list prices, not measurements. Assume an agent workload that consumes 1M input tokens in total, with 80% hitting cache (repeated system prompts and trajectory prefixes are the norm for agents), and produces 200K output tokens.
| Model | Fresh input | Cached input | Output | Real total | Saved vs. list price |
|---|---|---|---|---|---|
| GPT-6 Sol | $0.40 | $0.16 | $2.00 | $2.56 | approx. 12% |
| GPT-6 Luna | $0.02 | $0.008 | $0.10 | $0.128 | approx. 11% |
| Claude Opus 5.5 | $0.80 | $0.16 | $4.00 | $4.96 | approx. 17% |
| Grok 4.7 | $0.40 | $1.60 (no discount) | $1.20 | $3.20 | 0% (no public discount) |
Three counterintuitive facts fall out of this table.
First, Sol and Grok 4.7 share the same $2 input price, yet at a high cache hit rate their actual input costs differ by 10x: Sol pays $0.16 for the cached portion, Grok pays full freight at $1.60. To be fair to Grok: no published discount does not mean no caching; it just means the savings cannot be verified, so we book the cached portion conservatively.
Second, although Opus 5.5 costs 2x Sol on paper, its cache read price of $0.20 ties Sol's, so the cached portion costs exactly the same in absolute terms. The entire gap sits on the output side: $20 versus $10.
Third, Luna's real total of $0.128 is roughly 1/20 of Sol's and 1/39 of Opus 5.5's. If your task is within Luna's range, there is no suspense on price.
One engineering detail worth real money: Sol and Luna let you change the reasoning-effort tier mid-conversation without breaking the cache, and support explicit breakpoints to control cache prefix boundaries. Traditionally, switching effort tiers invalidates the cache and forces the prefix to be re-paid at full price. Mid-conversation re-tiering that preserves cache turns "save where you can, spend where you must" into an architectural feature. For the full methodology on cache economics, see our dedicated agent cache cost review (/en/posts/agentic-cache-cost-comparison-review); we will not repeat it here.
A caveat before moving on: the 80% hit rate above is an assumption, and real bills are extremely sensitive to it. Common cache-killers in agent stacks include edited system prompts, request gaps longer than the cache TTL, and load balancers scattering requests across instances. If your hit rate falls from 80% to 50%, Sol's cached-input cost climbs from $0.16 to $0.40. Replay real traffic before you commit, measure your actual hit rate, and recompute this table — with a different hit rate, the same rate card can support a different conclusion.
4. Benchmarks next to prices: what you can and cannot conclude
Ground rule first: every score below is an official vendor figure, produced with each vendor's own harness on its own task sets. We have not re-run anything. Comparisons across vendors and across benchmarks cannot support firm conclusions. We do only two things: place numbers side by side where multiple vendors published on the same benchmark, and mark "not published" everywhere else.
| Model | DeepSWE v1.1 | Terminal-Bench 4.0 | AutomationBench | OSWorld 2.0 |
|---|---|---|---|---|
| GPT-6 Sol | 68.8% (max) | Not published | 33.2% (xhigh) | 60.5% (xhigh) |
| GPT-6 Luna | 66.6% (max) | Not published | Absolute value not published | Absolute value not published |
| Claude Opus 5.5 | Not published | 66.4% | 40.0% | 81.8% |
| Grok 4.7 | 71.0% | 38.0% | Not published | Not published |
Two cells in this table deserve plain honesty.
First, Terminal-Bench 4.0: only Opus 5.5 (66.4%, officially described as leading GPT-6 Astra by 8.5 points) and Grok 4.7 (38.0%) published scores. Sol and Luna did not. We will not fill in guesses — no publication means no publication.
Second, DeepSWE v1.1: Grok 71.0% > Sol 68.8% > Luna 66.6%, with the most expensive model, Opus 5.5, absent from this benchmark and the cheapest of the three, Grok, on top (per each vendor's official figures). This ordering is worth chewing on. Converted to money per point, Grok has the lowest DeepSWE cost in this comparison — a $2/$6 rate card with 71.0% on a coding benchmark buys more points per dollar than Sol or Luna. But the same logic inverts on Terminal-Bench 4.0, where Grok's 38.0% sits far below Opus 5.5's 66.4%. One model, two benchmarks, two stories — that is precisely what "different harnesses cannot be compared" means in practice.
Opus 5.5's official materials add several numbers worth recording: 1846 Elo on GDPval-AA v2.1, 81.8% on OSWorld 2.0, 67.7% on HLE with tools, 54.4% on FrontierCode; officially, typical task total cost runs 40% below Opus 5 with 30%+ faster output. Sol's official figures: 33.2% (xhigh) on AutomationBench, above Claude Opus 5's best (max 26.9%) at roughly 9% of its per-task cost; 56.4% (max) on Agents' Last Exam across 55 industries at roughly 60% lower cost; 68.8% on DeepSWE, 1.1 points behind Claude Fable 5's best at about 1/5 the cost; 60.5% (xhigh) on OSWorld 2.0, comparable to Opus 5 at medium effort, again around 1/5 the cost. Luna: 66.6% on DeepSWE, close to Opus 5 at medium effort for 93% less cost; AutomationBench at high effort up 5.4 points over the previous Luna generation with costs down 58%.
Stack prices onto scores and the money-per-capability picture looks roughly like this. Opus 5.5 charges twice Sol's price for clear leads across several benchmarks — its bill makes sense when peak capability is a hard requirement. Sol is the textbook sweet-spot second flagship: distilled capability, half price, 90%-off cache reads, a steep cost-per-point curve. Luna pushes flagship-grade entry pricing down to a penny per 100K input tokens. Grok 4.7 posts the lowest cost per point on a coding benchmark in this field, but its Terminal-Bench gap warns you it is not an all-rounder. DeepSeek V4.1 Flash's 74.2% on DeepSWE is the highest coding score quoted in this article, and paired with time-of-use pricing its unit cost is extremely low for schedulable batch work.
5. Conclusions by scenario: four budgets, four bills
No rankings here — match your scenario to a row.
Peak capability, budget available: Opus 5.5. $4/$20 is not cheap, but against Opus 5 the vendor claims 40% lower typical task cost, and its $0.20 cache read ties Sol's. When one percentage point decides success or failure, double the price buys certainty.
Mainline workloads under cost control: Sol. $2/$10 with 90%-off cache reads, and official figures show multiple agent benchmarks matched or beaten at roughly 1/5 of last-generation flagship cost. For most teams this is the most balanced money-per-capability tier.
High-volume light tasks, absolute cost priority: Luna. At $0.10/$0.50, "per million tokens" almost feels like the wrong unit, and the cache discount suits high-frequency, repeated-prefix pipelines. The only homework is confirming your task difficulty sits inside its envelope — 66.6% on DeepSWE is close to, not at, Opus 5's mid-to-high intensity.
Output-heavy and coding agents: Grok 4.7 deserves a line in the ledger. The $6 output rate is the lowest here, and 71.0% on DeepSWE tops the three-way comparison — provided you accept the 38.0% on Terminal-Bench and treat its cache benefit as unknown.
Schedulable batch work: DeepSeek V4.1 Flash's time-of-use pricing is a genuine edge. Run batches off-peak at 1 CNY input and 4 CNY output — roughly $0.14/$0.57 (conversion cross-check in FAQ) — undercutting even Luna's flat price, if you can genuinely move work off-peak and accept the envelope implied by its 74.2% DeepSWE score.
Repeat the boundary once more: everything above is official vendor figures, untested by us; prices change, so check the official pricing pages before you commit. For "which model is actually smarter," go back to our GPT-6 Astra flagship review (/en/posts/gpt-6-astra-flagship-review). For the broader cost-performance picture, see our LLM cost-performance review (/en/posts/ai-llm-cost-performance-comparison-review) and the agent long-context cost review (/en/posts/agent-long-context-cost-review). For the context behind Opus 5.5's launch, read our Claude Opus 5 hotspot (/en/posts/claude-opus-5-hotspot).
FAQ
Q1: What does the 90% cache discount on Sol and Luna mean?
A1: Cached input reads bill at 10% of the list input price: roughly $0.20 per million tokens for Sol and $0.01 for Luna. Only "reads" are discounted; first writes and uncached fresh input still bill at full price. Both models also support mid-conversation effort re-tiering without breaking the cache, plus explicit breakpoints to control cache prefix boundaries.
Q2: Opus 5.5's list price is 2x Sol's — is the real bill also 2x?
A2: Not necessarily. Under this article's worked example (80% cache hits), Opus 5.5 totals $4.96 versus Sol's $2.56, about 1.9x. The cached portions tie at $0.16 each; nearly the whole gap sits on output. The higher your hit rate, the closer the real gap gets to 2x; with low hit rates the input-side gap narrows.
Q3: Does Grok 4.7 really have no cache discount?
A3: The precise statement is "no published cache discount data." Our source material contains no cache pricing for it, so the real-bill table books its cached portion at full price, conservatively. That does not mean the caching mechanism is absent — only that the savings cannot be verified.
Q4: How do you compare DeepSeek V4.1 Flash's time-of-use pricing with dollar rates?
A4: A direct comparison would be sloppy. DeepSeek charges 1 CNY off-peak / 2 CNY at peak for input, 4 CNY / 8 CNY for output. Using $0.10 = approx. 0.7 CNY (a conversion relayed by AI-tool aggregators), that is roughly $0.14/$0.57 to $0.29/$1.14. The two currencies are different pricing systems, so treat this as an order-of-magnitude reference only.
Q5: Grok is cheapest yet tops DeepSWE — best value for money, then?
A5: On that one benchmark (Grok 71.0% > Sol 68.8% > Luna 66.6%, per official figures), Grok's cost per point is indeed the lowest. But the same model posts 38.0% on Terminal-Bench 4.0, well below Opus 5.5's 66.4%. Different benchmarks with different harnesses cannot settle conclusions; choose based on which benchmark your actual workload resembles.