Week 37 pricing briefing

Cheaper Flash tokens. A new voice-session bill.

September 7-13, 2026. DeepSeek lowers Flash API rates, OpenAI prices GPT-Live by the minute, and restricted research-model billing gets a future start date.

Prepared September 13, 2026. Coverage through 01:05 UTC; Sunday is still in progress.

Your Wallet This Week

#1DeepSeek Flash: 100 million uncached input tokens off-peak now cost $15 instead of $22. Same volume, $7 less.
#2GPT-Live 1: the new voice-session charge is $0.05/minute, billed per second. Backend model and tool usage cost extra.
#3Rosalind research: announced input/output rates are $5/$25 per million tokens. Billing starts October 5, not this week, and access is restricted.

Price Summary

Usage	Before	Current / announced	Impact	Who It Affects
Flash cache-hit input	$0.007 / million	$0.003 / million	Lower rate	Cached prompts, off-peak
Flash cache-miss input	$0.22 / million	$0.15 / million	Lower rate	New prompt tokens, off-peak
Flash output	$0.66 / million	$0.60 / million	Lower rate	Output tokens, off-peak
GPT-Live 1	New GA launch price	$0.05 / minute + backend and tools	New session fee	Voice-agent developers
GPT-Rosalind	Prior price not established	$5 input / $0.50 cached / $25 output per million	Billing October 5	Approved life-sciences researchers

USD. Flash rows compare retired V4-Flash with V4.1-Flash at off-peak rates per million tokens; peak rates are double. GPT-Live is billed per second with backend/model/tool charges separate. Rosalind is a future billing announcement, not an already-effective price increase. No historical $0 price is assumed.

Previous report: Week 36, August 31-September 6, 2026. Browse the full briefing archive.

Flash gets cheaper. Pro is staying available.

DeepSeek released V4.1-Flash on September 10 and put its new prices into effect at 04:00 UTC that day. Use deepseek-flash. The old V4-Flash and V4-Flash-Vision-Exp models are retired; their old IDs temporarily route to V4.1-Flash.

Off-peak uncached input falls from $0.22 to $0.15 per million tokens, cache-hit input from $0.007 to $0.003, and output from $0.66 to $0.60. The before figures come from DeepSeek's official August pricing announcement, not an inferred historical price.

For 100 million uncached input tokens and 10 million output tokens, the off-peak bill changes from $28.60 to $21.00. This assumes the same token volumes; a replacement model may consume different amounts on the same task.

Peak hours are Monday-Friday, 01:00-04:00 and 06:00-10:00 UTC, excluding Chinese public holidays. All other hours, including weekends and Chinese public holidays in full, are off-peak. These are API charges, not a chatbot subscription price.

Important correction to the launch notice: the September 10 release announced a Pro-to-Flash routing change for September 14. The current pricing documentation now explicitly says DeepSeek will continue providing V4 Pro after that date with billing unchanged. Do not migrate solely because of the older deadline. The reversal footnote itself is undated, so we are not presenting it as a separately dated news event.

Who This Affects

Direct DeepSeek API users running agents, batch jobs or repeated prompts. Rebudget Flash by cached input, uncached input and output separately. Benchmark the replacement. Pro users should follow the current continuation notice, not the superseded launch plan.

Official sources: DeepSeek's dated September 10 release supports the new rates and transition; its dated August 13 announcement preserves the prior rates; the current Models & Pricing documentation corroborates current rates and Pro continuation. Current pricing alone is not used to prove the historical change.

GPT-Live starts at five cents a minute, not five cents all-in

OpenAI's September 10 API changelog announced GPT-Live 1 general availability at $0.05 per voice-session minute. It is billed per second, without rounding each session up to a full minute.

That is the session charge. Backend model and tool usage are charged separately. A budget that multiplies minutes by five cents and stops there is incomplete.

This is a verified launch price, not a demonstrated price cut. We have not established an earlier comparable per-minute rate, and token-priced Realtime usage is not an interchangeable denominator.

Who This Affects

Developers building voice agents. Measure session duration, backend model usage and tool charges separately before estimating cost per call. Do not treat $0.05/minute as an all-inclusive service price.

Official sources: OpenAI's dated September 10 API changelog establishes the launch price; the current GPT-Live pricing table corroborates the current rate. No historical reduction or subscription saving is claimed.

Research teams get an October billing date

On September 8, OpenAI announced GPT-Rosalind general availability through its trusted-access program for approved internal life-sciences research. This is not unrestricted public access.

Announced standard rates are $5 per million input tokens, $0.50 per million cached input tokens, and $25 per million output tokens. Billing begins October 5, 2026, not during Week 37.

No historical price has been established here. Treat this as a future-budget notice, not a claim that a free public model just became paid.

Who This Affects

Approved research teams using GPT-Rosalind. Confirm eligibility and model access, then budget the separate token categories before October 5. Most general-purpose chatbot subscribers are unaffected.

Official source: OpenAI's dated September 8 API changelog establishes eligibility, announced rates, and the future billing date. Checked September 13; no historical price is inferred.

Switch Watch

Act on the billing conditions, not just the headline.

Rebudget Flash

Test the cheaper replacement

Same-volume off-peak costs are lower. Benchmark task quality and token consumption before assuming the same reduction in your total bill.

Price the full call

Keep backend costs visible

For GPT-Live, add model and tool charges to voice-session time. No cross-vendor saving is established by the five-cent headline alone.

No rushed migration

Keep Pro's latest notice

Current DeepSeek documentation says Pro continues after September 14 with unchanged billing. The earlier forced-routing announcement is superseded.

Scope: one verified before/after rate cut, one launch-price announcement and one future billing notice. We do not claim these were the only market changes. Routine corrections from the wider price audit are not relabeled as Week 37 news. All sources are official; general AI news and unsupported savings claims are excluded.