Five frontier labs updated their flagship lineups in the same news cycle, and the standout signal isn’t a benchmark chart — it’s price convergence. Moonshot’s Kimi K3, xAI’s Grok 4.5, fresh Gemini Flash variants from Google, and tracker listings for Claude Sonnet 5 and GPT-5.6 all surfaced in recent weekly coverage, with per-token pricing now clustering within a few dollars of each other across vendors.
Why it matters
For builders, the last 18 months of frontier-model competition were mostly about raw capability leapfrogging. This wave looks different: labs are converging on similar price bands ($1.50–$5 input, $6–$30 output per 1M tokens) while pushing context windows toward the 1M-token mark. That means the deciding factor for most production apps is shifting from “which model is smartest” to “which model is cheapest at the context length and latency I actually need.”
What changed vs the previous generation
- Google: Added Gemini 3.6 Flash and 3.5 Flash Lite variants, extending the speed-optimized tier rather than replacing the flagship Pro line — a bet on cheap, fast inference for high-volume agentic use cases.[1]
- xAI: Grok 4.5 is listed as generally available with a 500K context window and $2/$6 pricing — competitive on cost but now the smallest context window among the majors compared here, a gap versus rivals pushing past 1M tokens.[2]
- Anthropic: Claude Sonnet 5 appears with introductory pricing of $2/$10 per 1M tokens, alongside a tracked “Fable 5” variant — introductory pricing signals Anthropic is still testing price elasticity before settling on standard rates.[2]
- OpenAI: GPT-5.6 Sol is tracked as GA with a 1.05M context window and the highest price point in this set at $5/$30 — the widest input/output price spread, suggesting heavier output-token cost engineering.[2]
- Moonshot: Kimi K3 is named in weekly roundups alongside the above, though detailed pricing and context specs weren’t available in this briefing — a gap worth tracking as more detail surfaces.[1]
Pricing and specs compared
| Model | Status | Context window | Input $/1M tokens | Output $/1M tokens |
|---|---|---|---|---|
| Gemini 3.5 Flash | GA | Not specified | $1.50 | $9 |
| Grok 4.5 | GA | 500K | $2 | $6 |
| Claude Sonnet 5 | GA (introductory pricing) | Not specified | $2 | $10 |
| GPT-5.6 Sol | GA (Jul 9) | 1.05M | $5 | $30 |
Figures are drawn from a secondary model-tracker summary, not primary vendor pricing pages — treat exact numbers as directional rather than final, and verify against each provider’s official docs before budgeting a production workload.
Trending themes
- Price compression at the low end: Flash- and speed-tier models are landing in the $1.50–$2 input range, making high-volume agentic workloads meaningfully cheaper than a year ago.[2]
- Context window as differentiator: GPT-5.6 Sol’s 1.05M window stands out versus Grok 4.5’s 500K, and it’s now a marketing point as much as chat quality is.[2]
- Agentic and coding benchmarks over chat polish: Grok, Claude, and Gemini are all being pitched primarily on tool-use and coding performance rather than conversational quality alone.[1][2]
The honest caveat
Some of the model names and version numbers here — GPT-5.6, Gemini 3.6 Flash, Claude “Fable 5” — come from a secondary tracker rather than confirmed vendor announcements in this briefing, and naming conventions across trackers can drift from official releases. We’re reporting what the sources say, but builders should cross-check exact model IDs and current pricing directly with OpenAI, Google, Anthropic, and xAI before wiring anything into production billing.
What it means for builders
If you’re shipping a product right now, the practical takeaway isn’t “switch to the newest model” — it’s that per-token economics have shifted enough to revisit your model routing. Latency-sensitive, high-volume features (autocomplete, classification, cheap agent loops) increasingly belong on Flash-tier or Grok-tier pricing, while long-document or long-context tasks may justify GPT-5.6 Sol’s premium given its 1.05M window. Run your own eval on your actual prompts and output lengths — output-token pricing spreads (as wide as $6 to $30 per 1M here) can flip the cost calculus depending on how verbose your use case is.