
The short version: GPT-6.2 slashes official prices by 20%, which is genuinely good news for buyers - yet platform discounts still run far deeper. As of November 18, 2026, the current gpt-6-astra listing trades at $0.099 per million input tokens on the deepest discount, under 1% of list price. In the same week Alibaba's Qwen4-Max and Tencent's Hy4 flash pushed per-task cost and cache pricing to new lows. Should you wait for GPT-6.2? If the budget rules, stay on today's cheap tiers. If you need the newest capability, wait for listings and compare effective prices. If you run high-volume traffic, look at cache pricing first.
On November 17 local time, OpenAI shipped GPT-6.2 Astra and GPT-6.2 Sol. Astra targets hard professional work at $8 per million input tokens, $40 output and $0.40 cached input - 20% below the previous GPT-6 Astra on input and output and 60% below on cached input - with roughly 1.05M tokens of context. Sol stays the everyday model at the previous $2 and $10.
The day before, Alibaba released Qwen4-Max: 2.4T total parameters, roughly 71B active per token, native vision and 1M-token context. Official figures put average per-task cost about 60% under Qwen3.8-Max, with cache hits as low as 0.5 yuan per million tokens and Terminal-Bench 4.0 token use down from 289K to 151K. Tencent's Hunyuan launched Hy4 flash the same day: 224B total, 12B active, multimodal, 1M context, at 0.5 yuan input, 2 yuan output and 0.02 yuan for cache hits - about 1/20 the per-task cost of Hy4 proper. Figures are from vendor announcements and press reports.
All three point the same way: competition has moved from raw capability to cost per completed task, with cache pricing and token efficiency in the spotlight.
A 20% cut only means something against the base. GPT-6.2 Astra still lists at $8 per million input tokens - squarely flagship territory - while the current GPT-6 Astra on the marketplace already sells below 1% of list. The cross-check is neat: our ledger prices gpt-6-astra at $10 input, $50 output and $1 cache read, exactly the base the 20% and 60% cuts are measured against. The cut is real, and it is measured from the old list.
| Model | List (in/out) | Deepest discount | Effective (in/out) | Sellers |
|---|---|---|---|---|
| gpt-6-astra | $10 / $50 | 0.99% of list | $0.099 / $0.495 | 10 (3 certified) |
| gpt-6-sol | $2 / $10 | 1.65% of list | $0.033 / $0.165 | 8 (3 certified) |
| gpt-6.1-sol | $2 / $10 | 1.98% of list | $0.0396 / $0.198 | 8 (2 certified) |
| gpt-5.6-sol | $4 / $20 | 2.2% of list | $0.088 / $0.44 | 8 (2 certified) |
These listings carry real traffic. Measured on-platform as of November 18, 2026: gpt-6-astra logged close to 900 calls over the past 24 hours at a 99.9% success rate, with a P50 first-token latency around 10 seconds and P50 output speed near 15 tokens per second. Ten sellers supply it. The price is not a shelf ornament.
Qwen4-Max competes on token efficiency rather than leaderboard rank: about 60% lower cost per task than Qwen3.8-Max, cache hits at 0.5 yuan per million tokens. The previous generation qwen3.8-max is listed here at $1.79 / $5.38 official, $0.22 cache read, and $0.896 / $2.69 effective on the deepest 50% discount (data as of November 18, 2026). Qwen4-Max itself is not on the plaza yet; when it lands, compare on the same terms.
Tencent's Hy4 flash is even more aggressive: 0.5 yuan input, 2 yuan output, 0.02 yuan cache hits - roughly 1/20 the per-task cost of Hy4 proper by the vendor's account. What we list today is hy4-preview, a different SKU in the same family: $0.896 / $2.69 official, $0.045 cache read, $0.246 / $0.74 effective at the deepest 27.5% discount. If flash arrives at similar terms, the marginal cost of routine work drops again.
Three situations, three answers. Budget-sensitive, routine work: switch to today's cheap tiers - gpt-6-sol at $0.033 per million input tokens handles writing, summaries and multi-step pipelines, two orders of magnitude under GPT-6.2 Sol's list price. Newest capability for long-running agents or brutal refactors: the official cut is a good sign, but the smart move is to wait for GPT-6.2 listings and compare seller discounts - on the gpt-6-astra precedent (0.99% of list), deep discounts tend to appear early. High-volume work with repetitive prefixes: treat cache pricing as a hard requirement. Qwen4-Max at 0.5 yuan and Hy4 flash at 0.02 yuan per cache hit both signal that repeated calls are nearly free.
The chart (data as of November 18, 2026) puts both forces on one axis: GPT-6.2's list price is 20% below its predecessor, while deepest seller discounts on current models run 72.5% to 99.01% off list. The official cut is visible; the seller discount is what actually moves the bill.
One, cut cost today: sort by effective input price on the model plaza and move routine work to gpt-6-sol or gpt-6.1-sol, both under 2% of list. Two, wait for GPT-6.2 the right way: validate your workflow on current Astra with light traffic, then compare effective prices on the same prompt once the new version is listed - never switch on list price alone. Three, build caching into high-volume jobs: fixed prefix, one warm-up call, then watch the cache-read share of your bill. It pays back faster than swapping models. The full discount rules live on the pricing page.
Prompt caching stores a repeated prompt prefix and bills cache reads at 5-12% of normal input. Real prices, hit conditions, and how to cut your API bi
Twelve Claude versions are on sale. The newest opus-5-5 costs $8.91 a month; the oldest opus-4-6 costs $49.50. Full table and a pick for each workload
gpt-5.6-terra and deepseek-v4.1-flash win on price and reliability; gemini-3.8-flash is fastest. Effective prices, measured speed, and a three-tier pi
Point your base URL at the platform and one key covers all 138 on-sale models across 51 vendors, in both OpenAI and Anthropic request formats. Copy-pa
Hailuo 2.3-fast has the lowest effective price (about $0.00185/s), Kling 3.0 offers the fullest range up to 4K, and Seedance 2.5 costs most but render