DeepSeek V4 Flash, GPT-5.6 Luna, or Terra for fiction? Put the two test suites on the right footing first
A data-backed comparison of DeepSeek V4 Flash 0731 and GPT-5.6 Luna/Terra for fiction: official input, cache, and output rates; 32K long-form completion; time to first prose; 450-700-character compliance; reasoning exhaustion; agent tools; and cache reads. The two suites were not one head-to-head blind test: Terra medium won the Luna/Terra matrix, while DeepSeek none is the cheapest practical tier.
Read the three result pages in isolation and you can get three answers that sound incompatible. Terra medium won the formal Luna/Terra blind review. DeepSeek V4 Flash is dramatically cheaper. Luna high is the most practical quality tier inside Luna. What was missing was not another synthetic super-score, but a page that separates matched evidencefrom cross-suite buying decisions.
This comparison uses two real API suites completed on July 31, 2026. It does not add synthetic samples or treat endpoint completion as fiction quality. DeepSeek traffic went directly to the official API; the formal Luna/Terra fiction matrix used Requesty Responses. Both suites covered 32K output, consecutive continuation, reasoning, caching, and agents, but their stories, chain counts, and review instruments differed.
What can be compared across the two suites?
| Dimension | DeepSeek V4 Flash 0731 | GPT-5.6 Luna / Terra | Can it share one ranking? |
|---|---|---|---|
| Production route | Official DeepSeek Chat / Responses | Requesty Responses; separate dual-relay production verification | Observed behavior only |
| Long-form scale | One archived 20-round chain; one 12-round chain per original effort | 28 planned chains up to 20 rounds; 548 complete prose calls | No shared success rate |
| 32K behavior | High/max fiction 12/12 each; complex max cards still zeroed out | All 548 completed calls had prose; five streams incomplete | Truncation boundaries are comparable |
| Length target | Original chain requested 450-700 Chinese characters | Both original fixtures requested 450-700 characters | Compliance is informative, fixtures differ |
| Quality review | Anonymous high/max A/B; max narrowly won both reviews | Two isolated model sessions, five dimensions, 84 rows each | Scores cannot be converted |
| Agents | Auto completed two steps; required + max returned 400 | Both relays completed two-step Luna/Terra tool loops | Parameter boundaries are comparable |
| Caching | Read hits; write tokens null | Read hits on both relays; write tokens null | Same conclusion |
One pricing ruler: DeepSeek is cheapest, Terra is most expensive
These are the official short-tier dollar rates on July 31, 2026, followed by Foreverse's uniform 50% markup in credits. One credit equals $0.0001. DeepSeek does not quote a separate cache-write tier. Luna/Terra cache writes are billed only when the upstream supplies authoritative cache_write_tokens; null remains unknown.
| Model | Official uncached input $/M | Official cache read $/M | Official cache write $/M | Official output $/M | App +50% credits/M: input / read / write / output |
|---|---|---|---|---|---|
| DeepSeek V4 Flash | $0.14 | $0.0028 | Not separately quoted | $0.28 | 2,100 / 42 / — / 4,200 |
| GPT-5.6 Luna | $0.20 | $0.02 | $0.25 | $1.20 | 3,000 / 300 / 3,750 / 18,000 |
| GPT-5.6 Terra | $2.00 | $0.20 | $2.50 | $12.00 | 30,000 / 3,000 / 37,500 / 180,000 |
Luna's uncached-input rate is about 1.43 times DeepSeek's, its cache-read rate 7.14 times, and its output rate 4.29 times. Terra's output rate is about 42.86 times DeepSeek's. The uniform markup preserves those ratios. Sources: DeepSeek official pricing and OpenAI official pricing.
Observed long-form runs: divide by completed rounds, but do not pretend the prompts match
A 12-call total and an 80-call total are not comparable, so the table shows observed average cost per completed round. DeepSeek's rows multiply official cost by 1.5 before division and omit per-request ceiling effects. Luna/Terra rows use the formal matrix's per-request normalized user prices. This is still not a strict price benchmark because input length, story, and cache rate differ.
| Setting and envelope | Prose completed | 450-700 chars | Median first prose | Observed average / complete round | Quality evidence |
|---|---|---|---|---|---|
| DeepSeek none · 8K | 12/12 | 0/12 | 0.75s | about $0.000580 | No three-way matched score |
| DeepSeek low · 8K | 12/12 | 0/12 | 7.06s | about $0.000931 | No visible gain over none in this chain |
| DeepSeek high · 32K | 12/12 | 3/12 | 89.5s | about $0.003440 | Lost the anonymous A/B to max |
| DeepSeek max · 32K | 12/12 | 2/12 | 43.5s | about $0.003198 | Narrowly won both anonymous A/B reviews |
| Luna none · 32K | 80/80 | 21/80 | 2.256s | $0.006008 | Five-dimension mean 3.658 |
| Luna high · 32K | 77/78 | 32/77 | 16.420s | $0.008469 | Mean 4.175; best Luna setting |
| Luna max · 32K | 80/80 | 27/80 | 14.652s | $0.008514 | Five-dimension mean 3.950 |
| Terra medium · 32K | 78/79 | 34/78 | 3.429s | $0.057617 | Mean 4.725; winner of that matrix |
Three conclusions survive the protocol difference. DeepSeek is the decisive cost tier, especially at none. Luna high produced a real quality increase inside its matched matrix. Terra medium's large premium bought the strongest and most stable quality evidence in that suite. The table does not prove that cheap DeepSeek must write worse, nor that Terra's 4.725 has already beaten DeepSeek head to head.
32K helps all three; complex DeepSeek max still exposes the sharpest reasoning risk
Under an old 8K envelope, DeepSeek high spent 8,191 output tokens entirely on reasoning and returned no prose. Luna's earlier screen showed the same failure shape at 8,192. At 32K, DeepSeek high and max both completed 12-round ordinary fiction chains, while all 548 completed Luna/Terra calls contained prose. A global 8K cap was clearly too tight.
The complex-task result is more cautionary. DeepSeek high completed four genre cards but waited 236 seconds for visible text and used 27,099 reasoning tokens. Max waited 293.9 seconds, spent all 32,767 tokens on reasoning, and returned zero prose. This Luna/Terra suite did not run a matched formal card-authoring exam, so it cannot establish that either writes cards better. It does establish that production must still detect incomplete/length + zero prose rather than treating 32K as a guarantee.
Agents and caching: all are usable, but one forced tool policy does not fit all
| Capability | DeepSeek V4 Flash | Luna / Terra | Application policy |
|---|---|---|---|
| Streaming text | Official Chat and Responses completed | Inroi Chat and Requesty Responses completed | Keep one app-facing Chat SSE interface |
| First tool turn | Responses auto emitted a function call | Both relays emitted tool calls | Convert centrally by upstream protocol |
| Tool-result follow-up | Second request returned 200 and the right answer | Both models and routes closed the loop | Settle final usage once |
| Forced tools | thinking=max + required returned 400 | Named selection worked after conversion | Use auto plus an explicit prompt for DeepSeek thinking |
| Cache reads | 512-token hit on the second agent turn; content hits too | Hits observed through both relays | Bill authoritative cached tokens only |
| Cache writes | cache_write_tokens=null | cache_write_tokens=null | Unknown is not zero |
What to choose: four useful tiers, not one invented champion
Lowest-cost drafts and high-volume experimentation: DeepSeek none. It reached first prose in 0.75 seconds in its chain and costs the least, but missed the requested length in all 12 rounds; the application needs a firmer stop policy.
Routine passage continuation: Luna none/low or DeepSeek none. Choose DeepSeek for budget and Luna for a larger formal long-run evidence base. Neither family needs high effort as a global default.
A quality step-up for an important chapter: Luna high. It was the only Luna tier with a clear quality separation, at a 16.420-second median to prose. Xhigh and max did not improve the score further.
Best current quality evidence when budget permits: Terra medium. It scored 4.725 across two fixtures and four planned chains. That means “strongest current evidence,” not “already beat DeepSeek in the same arena.”
A true three-way winner requires a new paired exam: the same original fixtures, context, 32K envelope, rounds, and blind packs, with DeepSeek none/low, Luna high, and Terra medium all present. Until then, this page is enough to choose on price, engineering risk, and application defaults without laundering cross-suite numbers into a fake leaderboard.
Read the source result pages: the DeepSeek V4 Flash 0731 release benchmark, the 548-completion Luna/Terra blind review, and the Luna/Terra relay, cache, and pricing field test. The six unedited Chinese outputs put Luna high, Luna none, and Terra medium behind anonymous labels until the reveal.
FAQ
Which writes fiction best: DeepSeek V4 Flash, Luna, or Terra?
There is no defensible three-way champion yet because DeepSeek and Luna/Terra did not run the same fixtures, blind packs, or score sheet. The strongest current quality evidence is that Terra medium scored 4.725/5 and Luna high 4.175/5 in the 548-completion Luna/Terra matrix. In DeepSeek's separate 32K original chain, two anonymous reviewers narrowly preferred max over high, but those votes cannot be converted into the other leaderboard.
Is DeepSeek V4 Flash really cheaper than Luna?
At the official July 31, 2026 short-tier rates, DeepSeek charges $0.14 uncached input, $0.0028 cached input, and $0.28 output per million tokens; Luna charges $0.20, $0.02, and $1.20. Luna's output rate is therefore about 4.29 times DeepSeek's and its cache-read rate about 7.14 times. Foreverse applies the same 50% markup to both, so the ratios do not change.
Which model should handle routine passage-by-passage continuation?
Start with DeepSeek none when cost dominates; it was fastest and cheapest in its separate 12-round chain, but hit the 450-700-character target in 0/12 rounds, so it needs stronger length controls. Luna none/low has a larger formal evidence base for routine work. Raise a single important passage to Luna high. Choose Terra medium when the best quality evidence in the current matrix matters more than budget.
Does a 32K output cap eliminate empty prose for all three?
No. All 548 completed Luna/Terra calls contained prose, and DeepSeek high/max both completed their 12-round 32K fiction chains. But DeepSeek max still spent all 32,767 tokens on reasoning and produced zero visible text on a complex four-card task. A 32K envelope reduces truncation risk; it is not a capability guarantee.
Do DeepSeek, Luna, and Terra all support agent tools?
Yes, with different parameter boundaries. DeepSeek Responses completed a two-step tool loop with tool_choice=auto, while thinking=max plus required returned a real 400. Luna and Terra completed tool calls and tool-output follow-ups through Foreverse's Chat interface and backend Responses conversion on both relay routes. A production adapter cannot force every model into required mode.
Questions or ideas? Join our Discord →