Five Frontier Coding Models Shipped in 48 Hours. Here Is the Scoreboard.
It is Saturday, the launch dust has settled, and I finally have a full week of numbers to look at. Between the morning of July 8 and the evening of July 9, five frontier models aimed at coding and agentic work shipped into public availability. We covered the big ones one at a time as they landed. Now I want to put all of them on the same table, because the individual launch posts miss the thing that actually changed this week.
The leaderboard did not move. Anthropic still sits on top of SWE-Bench Pro. What moved is the floor: the cost of getting most of the way to the frontier fell off a cliff, and it fell in a single 48-hour window.
Who Shipped, and When
Here is the raw sequence, because the compression is the story. Five launches, two calendar days.
| Date | Model | Lab | Pitch |
|---|---|---|---|
| Jul 8 | Grok 4.5 | SpaceXAI | Cursor-trained coding model, cheap |
| Jul 9 | GPT-5.6 (Sol, Terra, Luna) | OpenAI | Three-tier family, new ChatGPT default |
| Jul 9 | Muse Spark 1.1 | Meta | Meta's first paid API model |
| Jul 9 | Seedream 5.0 Pro | ByteDance | Multilingual text-and-layout image model |
I am counting GPT-5.6 as one launch even though it is three separate models, because they shipped together and share a pricing philosophy. Add Anthropic's Sonnet 5 (June 30) and the July 1 return of Fable 5 and Mythos 5 from their export-control pull, and the buyable frontier expanded by roughly eight distinct models in eleven days. That is not a normal cadence. That is a scramble.
The Coding Scoreboard
Everyone leads their launch post with a benchmark. The problem is that no two labs run the same harness, so the numbers are not strictly comparable. I am going to show them anyway, because the pattern survives the noise. These are SWE-Bench Pro resolve rates as reported by each lab, which means they are vendor numbers, not neutral ones. Read them as a shape, not a ruling.
| Model | SWE-Bench Pro | Note |
|---|---|---|
| Claude Mythos 5 | 80.3% | Leads, vendor scaffolding |
| Claude Fable 5 | 80.0% | Score contested, replication underway |
| Grok 4.5 | 64.7% | Cursor-trained |
| GPT-5.6 Sol | 64.6% | OpenAI flagship tier |
| GPT-5.6 Terra | 63.4% | Mid tier |
| GPT-5.6 Luna | 62.7% | Cheapest tier |
| GPT-5.5 (April flagship) | 59.4% | For reference |
Look at the cluster. Grok 4.5, all three GPT-5.6 tiers, and last quarter's flagship land inside a six-point band. The two Claude models sit fifteen points clear of the pack. If you only read the leaderboard, you would conclude Anthropic won the week and everyone else traded rounding errors.
That conclusion is wrong, and the reason it is wrong is the second table.
The Number That Actually Moved: Price
The leaderboard measures capability. Nobody ships production on the leaderboard. They ship on a budget. So here is the same window priced out, per million tokens.
| Model | Input | Output |
|---|---|---|
| GPT-5.6 Luna | $1.00 | $6.00 |
| Grok 4.5 | $2.00 | $6.00 |
| Claude Sonnet 5 (intro) | $2.00 | $10.00 |
| GPT-5.6 Terra | $2.50 | $15.00 |
| GPT-5.6 Sol | $5.00 | $30.00 |
| Claude Fable 5 | $10.00 | $50.00 |
Now overlay the two tables. Grok 4.5 and GPT-5.6 Luna score within a couple of points of Sol, and within fifteen points of Fable 5, while costing one fifth to one tenth as much on output. The gap between the best model and the good-enough model has never been this cheap to skip.
The sharpest version of this shows up on DeepSWE, a long-horizon engineering benchmark, when you divide score by dollars. By the estimates circulating this week, Luna returns roughly 24 benchmark points per API dollar. Claude Opus 4.8 returns about 4.5. Fable 5 returns about 3.2. Luna is not the smartest model on that test. It is doing something like five to seven times the work per dollar, and that is the ratio a CFO signs off on, not the leaderboard rank.
Two Bets on the Same Table
What I find interesting is that these labs are not running the same play. They are running opposite ones, and both shipped in the same 48 hours.
Anthropic is defending the ceiling. Mythos 5 and Fable 5 hold a real, measurable lead on the hardest coding evals, and Anthropic is charging accordingly: $10 input and $50 output on Fable 5 is roughly ten times Luna on the way in. The bet is that a slice of the market will always pay a premium for the last fifteen points, because on genuinely hard problems those points are the whole job.
OpenAI and SpaceXAI are attacking the floor. Luna at a dollar in and Grok 4.5 at two dollars in are not trying to win the benchmark. They are trying to make the benchmark irrelevant to the purchase decision by getting close enough that price becomes the only variable left. SpaceXAI trained Grok 4.5 inside Cursor on real developer sessions, which is a different way of chasing the same goal: match the frontier on the work people actually do, then win on cost and distribution.
Both bets can be right at once, and I think they are. The market is splitting into a thin premium tier where Anthropic prices like a specialist, and a thick commodity tier where a dollar of output buys you most of the frontier. This week is the clearest snapshot yet of that split, because you can see both strategies land on the same two days.
The Caveat I Keep Repeating
Every number in the first table is vendor-reported. Fable 5's SWE-Bench Pro score is already contested, with independent evaluators noting it was produced on Anthropic's own scaffolding rather than a neutral harness. OpenAI's own comparison tables put Mythos 5 ahead of Sol on this exact benchmark, which is an unusual thing for a lab to publish about a competitor unless the gap is real. The Artificial Analysis Coding Agent Index, which pairs models with a fixed harness, tells a slightly different story and puts Sol at the top of the new entrants. Different harness, different winner. That is the whole problem with launch-week benchmarks.
So treat the ranks as provisional. The prices are not provisional. Prices are the one number a lab cannot fudge, and the price collapse is what I would build around if I were shipping this quarter. You can run your own workload against the current numbers on our cost calculator and track the live leaderboard on our benchmarks page.
What I Am Watching Next
Three things over the next two weeks. First, independent replication of the Claude SWE-Bench Pro numbers on a neutral harness, which will tell us whether the fifteen-point lead is real or scaffolding. Second, Gemini 3.5 Pro, which Google slipped from June into a July general-availability window and which has not entered this scoreboard yet. When it does, it lands straight into the commodity tier and the price pressure gets worse. Third, whether Anthropic's premium holds once buyers have a full month of production data on the cheap tier instead of a launch-day benchmark.
Our Take
The headline this week was five models in 48 hours. The actual event was quieter: the price of good enough dropped to a dollar, and the only labs still charging a premium are the two that can prove they earn it. Everyone else is now competing on the one axis they cannot spin.