DeepSeek V4-Flash vs GPT-5.6 Sol: $0.14 Against $5

DeepSeek opened the V4-Flash API in public beta on July 31, 2026, and announced it on X with an agent-capability claim plus two integration details: the model speaks the Responses API format natively and is "fully adapted for Codex." The launch post has since passed 8.7 million views. The model buyers are weighing it against for agent work is GPT-5.6 Sol, OpenAI's frontier tier, listed at $5.00 per million input tokens and $30.00 per million output since its July 9 launch and left at those figures by OpenAI's July 30 efficiency update, which cut Luna by 80% and Terra by 20%.

When we put Sol next to Qwen3.8-Max on August 3, only one side of that comparison had published token rates. V4-Flash's are now readable on Artificial Analysis's model page: $0.14 per million in, $0.28 out, and $0.003 per million on cache hits, all as listed there on August 2, 2026. The axis that decides this matchup sits underneath the benchmark tables, and it is this: what share of your agent's tokens are a repeated prefix, and whose servers hold that prefix.

"natively supports the Responses API format and is fully adapted for Codex"

DeepSeek put protocol support on the same line as the benchmark claim, which is a hint about how the two connect. Chat Completions, the endpoint DeepSeek has served for years, is stateless. Every turn your client resends the whole messages array: system prompt, all prior turns, every tool call, every tool result. A ten-step agent run transmits the step-one context ten times.

The Responses format inverts the arrangement. The server returns a response object with an id, the next turn sends only the new items plus a pointer back to that id, and DeepSeek reconstructs the rest of the history on its side. For a reasoning model there is a second effect worth more than the token accounting: the model's own reasoning items can survive a tool call instead of being thrown away, so after a tool returns, the model picks up a plan it already had.

Now the pricing consequence. A loop with a stable prefix turns most of its input into cache hits, and the cache-hit line is where DeepSeek's sheet gets strange. Artificial Analysis lists it at $0.003 per million tokens, a 98% discount off the $0.14 input rate, and ranks it first of the 101 models in its size class. The same page quotes a blended rate of $0.06 per million on a 7:2:1 mix of cache hits to fresh input to output, which is a fair description of an agent loop resubmitting a growing conversation.

Traffic with a different shape gets a different bill. One-shot classification over unrelated documents caches nothing and pays the full $0.14. Long answers pay $0.28, and V4-Flash is verbose: Artificial Analysis records 210 million output tokens spent getting through its nine evaluations, against a 100 million median for comparable models. Two other things follow from the Codex line. Migration cost drops, because Codex CLI and harnesses like Cursor can point at a custom provider without a chat shim. And your conversation state now lives on DeepSeek infrastructure, subject to whatever retention terms the platform publishes, which is a compliance question before it is an engineering one.

"We've massively upgraded its Agent capabilities"

Artificial Analysis scores the reasoning variant at maximum effort at 50 on its Intelligence Index, against a 25 median, and ranks it third of 101 large open-weights models. Three of the nine evaluations folded into that composite are agentic by the site's own labelling: GDPval-AA v2 on real-world work tasks, τ³-Banking on tool use, and Terminal-Bench v2.1 on terminal and coding work. The architecture underneath is 284B total parameters with 13B active, a 1M-token context window, and MIT-licensed weights on Hugging Face.

The envelope has hard edges. V4-Flash takes text and returns text, and Artificial Analysis states plainly that it is not multimodal, so a screenshot of a failing UI is outside what you can hand it. The Speed field on that page currently reads N/A, meaning nobody has published a throughput figure for it.

Sol's pitch runs on the opposite dimension. OpenAI's July 9 launch put serving throughput around 750 tokens per second via a Cerebras inference partnership, and pointed the tier at long-horizon agentic workflows, software engineering and cybersecurity. Its 88.8 on Terminal-Bench 2.1 is the figure that showed up in Alibaba's own comparison table on August 3, where Qwen3.8-Max came in 2.2 points behind at 86.6. Throughput is the part a user feels in an agent loop making dozens of sequential calls, and it is the part of Sol's price you are buying that a token rate does not describe.

"the best performance for its size"

The phrase belongs to Unsloth's model card for the 0731 GGUF builds, quoted by a user on the NVIDIA developer forum on July 31, alongside the claim that it is outperforming V4-Pro (Preview). Size is the honest frame for the whole comparison. Thirteen billion active parameters are answering against a frontier tier, and the gap in the price column is wider than the gap in any score column.

The multiples come out at roughly 36x on input and 107x on output. Anyone doing arithmetic on a monthly agent bill should hold one number in their head from that table: a million tokens of repeated agent context bills at about a third of a cent on the beta endpoint, and the same million tokens of fresh input on Sol bills at five dollars.

"placing it well above average among comparable models"

Read the qualifier in Artificial Analysis's sentence. "Comparable models" there means the 101 large open-weights models it tracks above 150B parameters, and Sol sits outside that class entirely. Buyers cross the class line anyway, so here is who lands on each side of it.

V4-Flash fits a team that owns its harness. You are writing the agent loop, you can keep a stable prefix so the cache rate applies, your work is text, and you are either comfortable sending source code through a China-hosted endpoint or prepared to serve the MIT weights yourself. Unsloth's card puts the full-precision Q8 build at 162GB, so self-hosting is a hardware decision before it is a licensing one. For a first paid API on a small budget, the beta endpoint is the one where an experiment costs pocket change.

Sol fits the opposite profile. Procurement has ruled on which vendors may see your code, latency is what your users experience, the hard steps in your pipeline justify a frontier tier, or your agent runs inside a product that picks the model for you. In that last case the comparison is academic, because the per-token difference is not yours to capture.

What's underappreciated here: the 98% cache discount is a property of your code as much as of the model. Harnesses that reshuffle tool definitions between turns, inject a timestamp near the top of the system prompt, or reorder retrieved file context will miss the cache on most calls, and the cheap model then arrives at the full $0.14 input rate with an output leg running about twice the median length. Choosing V4-Flash on the strength of its blended figure means committing to own your context management. Teams that would rather not think about prefix stability are buying a different product than the one on the price page.

Two dated events on the record already answer the question the benchmark tables leave open. On May 23, 2026, DeepSeek made its 75% cut on V4-Pro permanent, converting a promotion into a sticker price with no expiry, which is what a company does when it has measured its own unit economics at the low price and liked them. On July 30 OpenAI published an efficiency update that cut its two cheap tiers and left Sol at $5.00 and $30.00, the rates it launched with on July 9. Until one of those two sheets moves at the top, the 36x on the input line is the number you budget against.