July 14, 2026

GPT-5.6 Lands on Amazon Bedrock, and Claude Costs More in India Than the US

OpenAI's three-tier family went generally available on Amazon Bedrock on July 13 at the same per-token rates as the OpenAI API, with usage counting toward AWS commitments. Anthropic opened rupee pricing for Claude in India, its second-largest market, at prices above US levels once local tax is included. Plus OpenAI's own GPT-5.6 docs say leaner prompts cut API cost 33 to 67 percent.

GPT-5.6 Sol, Terra, and Luna are now generally available on Amazon Bedrock

GPT-5.6 went public on July 9 across ChatGPT, Codex, and the OpenAI API. Five days later it reached a second cloud. Amazon announced on July 13 that GPT-5.6 Sol, Terra, and Luna are generally available on Amazon Bedrock, the first non-OpenAI platform to host the family.

The pricing detail that matters for enterprise buyers: Bedrock charges the same per-token rates as the OpenAI API, and usage counts toward existing AWS commitments. Sol is $5/$30, Terra $2.50/$15, Luna $1/$6 per million input/output tokens. For teams already on a Bedrock commit or Enterprise Discount Program, that is a materially different bill than calling the OpenAI API directly, because the spend lands inside a commitment you have already negotiated instead of as a separate vendor invoice.

Regional availability is uneven. Sol, the flagship, is live in two AWS regions: US East (N. Virginia) and US East (Ohio). Terra and Luna add a third, US West (Oregon). Bedrock's inference engine pools capacity across regions and isolates each customer's throughput, and prompt caching carries over from the OpenAI side: explicit cache breakpoints, cached input at a 90 percent discount, and a 30-minute minimum cache life.

Bedrock adds a security layer OpenAI's own API does not. A zero-operator-access model enforced at the chip level means no AWS operator can read your prompts or completions. Every call runs under your IAM policies inside your VPC and logs to CloudTrail. One fine-print carryover: classifier-flagged traffic data is retained for up to 30 days for automated abuse detection, the same retention window that applies to Anthropic's Mythos-class models.

The announcement also surfaced benchmark numbers not in the original launch materials. On the Artificial Analysis Coding Agent Index, Sol scores 80, 2.8 points above the next-best model, while using less than half the output tokens, taking less than half the time, and costing about a third less. On ExploitBench, a cybersecurity research benchmark, Sol scores 73.5 percent versus 47.9 percent for GPT-5.5 at a comparable output-token budget. On Agents' Last Exam, Sol sets a new high of 53.6, outperforming the next-best model by 13.1 points.

Claude Pro, Max, and Team now sell in rupees in India, at prices above the US

Anthropic has started localizing Claude's subscription pricing in India, TechCrunch reported on July 13. India accounts for 5.8 percent of global Claude usage, making it the service's second-largest market after the US.

The rupee prices are higher than the US dollar prices, and they include local tax. On Claude's website in India, Pro is listed at 2,000 rupees a month when billed annually, about $21 versus $17 in the US. Max starts at 11,999 rupees a month, about $125 versus $100. Team plans start at 2,399 rupees per seat a month, about $25 versus $20. Prices on Claude's mobile apps vary slightly from the website. Anthropic has not yet enabled UPI, India's dominant instant payments network, so users pay by card or through Apple and Google app store billing. OpenAI added UPI support when it rolled out rupee pricing for ChatGPT last August.

The localization follows Anthropic's India buildout. A Bengaluru office opened in February, former Microsoft India head Irina Ghose was appointed to lead the country business in January, and partnerships with Infosys and Tata Consultancy Services were signed in recent months. The expansion hit a setback in June when Anthropic suspended Fable 5 and Mythos 5 access for non-US entities. Fable 5 access has since been restored, while Mythos 5 remains limited.

OpenAI's GPT-5.6 docs: leaner prompts cut cost 33 to 67 percent

The part of the GPT-5.6 general availability release that quietly changes your effective cost sits in the model guidance docs, not the pricing page. OpenAI tells developers to favor leaner prompts: remove repeated instructions, cut examples, and simplify tool descriptions. In a sample of internal coding-agent eval runs, configurations with leaner system prompts improved evaluation scores by roughly 10 to 15 percent while reducing total tokens by 41 to 66 percent and cost by 33 to 67 percent. OpenAI labels the ranges directional and tells teams to validate on their own workloads.

Two new API controls support the shift. A text.verbosity parameter lets you set a default detail level (low, medium, or high) across requests instead of stuffing brevity instructions into every prompt. And GPT-5.6 is more concise by default than GPT-5.5, so legacy "be concise" instructions can now make responses too short.

The reasoning ladder also grew. reasoning.effort now supports none, low, medium, high, xhigh, and max. OpenAI's migration advice is to start at your current GPT-5.5 or GPT-5.4 effort setting, then test the same setting and one level lower, since GPT-5.6 can often hold quality with fewer tokens. Pro mode, a Responses API execution mode that applies more model work to a single request, increases latency and aggregates the extra tokens into your usage at standard rates. It is a setting, not a separate model slug.

For a beat that tracks per-token prices, the implication is that the sticker rate understates how much GPT-5.6 actually costs to run. The same task may consume a third to two-thirds fewer tokens than it did on GPT-5.5 if you rework your prompts, and OpenAI is the one saying so.

Tracking

Gemini 3.5 Pro is widely reported to target July 17, three days out, and remains unconfirmed by Google. No model card, no API documentation, and no pricing table have been published. The public Gemini API still lists only gemini-3.5-flash and gemini-3.1-pro-preview. Analyst estimates put Pro at roughly $15/$60 per million tokens with a 2-million-token context window, but every specific number is provisional until Google posts. Gemini 3.5 Flash, the current Google frontier, is confirmed at $1.50/$9.

Claude Fable 5 included subscription access runs through July 19 at 11:59:59 PM PT, five days out. This is the third extension (the deadline has moved June 23, then July 7, then July 12, then July 19, each resolved at the wire). After July 19, every Fable token draws from usage credits at $10/$50, double Opus 4.8's $5/$25. Watch for a fourth extension around the 19th. Anthropic's support page makes no forward commitment beyond July 19.

DeepSeek V4 official launch is still targeted for mid-July with peak-hour pricing that doubles rates during Beijing business hours (9 to 12 and 14 to 18, which is 1 to 4 AM and 6 to 10 AM UTC). The legacy deepseek-chat and deepseek-reasoner endpoints hard-retire on July 24 at 15:59 UTC, 10 days out, routing to V4 Flash non-thinking and thinking modes until then.

OpenAI's July 23 deprecation wave is 9 days out. The official deprecations page lists gpt-5-chat-latest, gpt-5.1-chat-latest, five Codex variants, computer-use-preview, and deep-research models. The base GPT-5.4 model is not on the list, despite a blog claim to the contrary.

Claude Sonnet 5 introductory pricing of $2/$10 ends August 31. On September 1 it jumps to $3/$15. The newer tokenizer that Sonnet 5 shares with Fable 5 and Opus 4.7 and later produces about 30 percent more tokens for the same text, so the effective increase is closer to $3.90/$19.50, above Sonnet 4.6's $3/$15.

Mistral's frontier open-weight mixture-of-experts model is in early access with research, government, and industry partners, with general availability later this summer. No name, specs, benchmarks, or pricing have been confirmed. Mistral Large 3, the current flagship, is $0.50/$1.50 under Apache 2.0.

Current prices

All prices per 1M input/output tokens, verified against official pages on July 14, 2026.

GPT-5.6 Sol: $5 / $30. 1.05M context. Cache writes 1.25x, reads 90 percent off, 30-minute minimum. Prompts over 272K tokens billed at 2x input and 1.5x output.

GPT-5.6 Terra: $2.50 / $15. 1.05M context. Same cache terms.

GPT-5.6 Luna: $1 / $6. 1.05M context. Same cache terms.

Claude Opus 4.8: $5 / $25. Cache writes $6.25 (5m) or $10 (1h), hits $0.50.

Claude Sonnet 5: $2 / $10 through August 31, then $3 / $15 from September 1. Newer tokenizer adds about 30 percent more tokens, so effective post-September cost is closer to $3.90 / $19.50.

Claude Fable 5: $10 / $50. Cache writes $12.50 (5m) or $20 (1h), hits $1. Included in paid subscriptions through July 19, then credits-only.

Gemini 3.5 Flash: $1.50 / $9. 1M context. Batch $0.75 / $4.50. Gemini 3.5 Pro is not yet listed.

Grok 4.5: $2 / $6. 500K context. Not eligible for the 20 percent batch discount (grok-4.3 and grok-4.20 are). Priority processing is a 2x premium.

DeepSeek V4 Pro: $0.435 / $0.87 off-peak (cache miss). Cache hit $0.0036. 1M context. Peak hours double all rates at the mid-July launch.

DeepSeek V4 Flash: $0.14 / $0.28 off-peak. Cache hit $0.0028. 1M context.

Meta Muse Spark 1.1: $1.25 / $4.25. Cached input $0.15. 1M context, no long-context premium. Web search $2.50 per 1,000 queries. US-only waitlist.

Mistral Large 3: $0.50 / $1.50. Open-weight, Apache 2.0. Batch 50 percent off, cached input 90 percent off.